TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 24, No. 5, October 2026, pp. 1561 ∼ 1580 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v24i5.27678 ❒ 1561 NiteT r on: a no v el nighttime b unak en sea turtle species detection Muhamad Dwisnanto Putr o 1 , J ane Litouw 2 , Abdul Haris Onto wirjo 2 , Sherwin Reinaldo U . Sompie 2 , Hebr on Prasetya 1 1 Master Program of Informatics, Postgraduate Program, Sam Ratulangi Uni v ersity , Manado, Indonesia 2 Department of Electrical Engineering, Sam Ratulangi Uni v ersity , Manado, Indonesia Article Inf o Article history: Recei v ed No v 19, 2025 Re vised Jun 2, 2026 Accepted Jun 30, 2026 K eyw ords: Bunak en Computer vision Deep learning Nighttime Sea turtle Underw ater detection ABSTRA CT Automated monitoring is essential for sea turtle conserv ation, and vision-based algorithms of fer a non-intrusi v e approach to observing turtles without disturb- ing their natural habitat. Ho we v er , nighttime conditions introduce signicant challenges, as lo w illumination se v erely reduces the visibility of discriminati v e features. This limitation requires rob ust feature e xtraction to capture the lim- ited important features. Recent adv ances in computer vision, such as you only look once (Y OLO) 11-nano, enable real-time detection and allo w optimization under challenging conditions . In this study , we propose NiteT ron, a nighttime sea turtle species detector that uses a lightweight architecture with enhanced feature renement. The netw ork modies Y OLO11-nano to create a more ef- cient v ariant, Y OLO11-pico. T o impro v e detection in lo w-light scenarios, we introduce a strong feature modulation (SFM) module. This module enhances attention to rele v ant features by e xpanding the recepti v e eld in the nal back- bone stage. W e also incorporate a sparse separated con v olution (S2C) module to e xtract k e y information ef ciently while balancing accurac y and computational cost. The proposed mode l achie v es high precision with fe wer parameters and lo wer gig a oating-point operations per second (GFLOPs) than Y OLO11-nano. Experiments demonstrate superior performance, achie ving 0.682 mean a v erage precision (mAP)@0.5:0.95 for nighttime sea turtle detection. Inference tests also conrm ef cient central processing unit (CPU) deplo yment, reaching 21.18 frames per second (FPS) on a desktop and 2.65 FPS on an edge de vice. This is an open access article under the CC BY -SA license . Corresponding A uthor: Muhamad Dwisnanto Putro Master Program of Informatics, Postgraduate Program, Sam Ratulangi Uni v ersity Manado, Indonesia Email: dwisnantoputro@unsrat.ac.id 1. INTR ODUCTION Sea turtles play a crucial role in maintaining the st ability of marine ecosystems. Notably , Bunak en Marine P ark pro vides a natural habitat for three distinct sea turtle species and is recognized as a marine conser - v ation area with high biodi v ersity . The three species commonly found in Bunak en are the green, oli v e ridle y , and ha wksbill T urtle. Each of these species mak es a unique contrib ution to the m arine ecosystem. Therefore, a w areness of their location is crucial for protecting their habitat and population. T o address this issue, con- tinuous monitoring through automated detection systems po wered by computer vision is essential to pre v ent e xtinction and ensure long-term survi v al. These systems enable around-the-clock monitoring while minimizing J ournal homepage: http://telk omnika.uad.ac.id Evaluation Warning : The document was created with Spire.PDF for Python.
1562 ❒ ISSN: 1693-6930 resource and operational costs. A pre vious study de v eloped a you only look once (Y OLO)v7-based detection model combined with the bag of tricks-sort (BoT -SOR T) algorithm for automatic counting and tracking of green turtles in Japanese coastal areas [1]. The approach achie v ed promising results, with the best performance reaching a mean a v erage precision (mAP@50:95) of 0.620. Ho we v er , since the method relied on aerial drone vie ws, its ef fecti v eness w as limited when detecting underw ater turtles due to turbidity , surf ace ripples, and reduced visibility . On the other hand, the studies i n [2], [3] ha v e successfully implemented underw ater sea turtle detection as a foundational step to w ard automated conserv ation systems. Ho we v er , continuous monitor - ing f aces signicant challenges in nighttime en vironments due to lo w illumination. While pre vious studies on underw ater turtle detection ha v e sho wn promising results, the y are limited to daytime conditions, reducing their applicability in lo w-light scenarios. Lo w light conditions reduce k e y features use d to identify turtle species, such as shell patterns and coloration. Moreo v er , underw ater en vironments introduce additional challenges, such as motion blur and visual distortion, which further compromise recognition accurac y . In order to address these challenges, this research de v elops a model that performs accurately and ef ciently in nighttime conditions, making it suitable for deplo yment on lo w-cost de vices. The proposed model inte grates deep learning methods and ensures reliable detection and recognition of turtle species in lo w-light en vironments. This approach not only impro v es accurac y b ut also reduces reliance on human interv ention, minimizing errors and enhancing op- erational ef cienc y . Moreo v er , it enables continuous, real-time monitoring in challenging underw ater settings, supporting scalable and sustainable marine conserv ation ef forts. Con v olutional neural netw orks (CNNs) are widely used in object detection [4], [5]. One of the most well-kno wn object detectors is the Y OLO [6], which relies on CNNs as its core component and is capable of real-time operation. The demand for lightweight and ef cient algorithms in computer vision led to the de- v elopment of Y OLO11-nano [7]. These netw orks utilize fe wer parameters and gig a oating-point operations per second (GFLOPs) to operate. Ho we v er , it deli v ers lo wer precision compared to other Y OLO11 v ariants. Hence, the performance needs to be enhanced to impro v e its ability to handle nighttime and underw ater chal- lenges. Its lightweight and ef cient design demonstrates strong potential for further optimization under such demanding conditions. Although impro ving performance often demands higher computational costs , se v eral studies ha v e introduced lightweight modules that boost recognition accurac y without signicantly increasing comple xity [8]-[10]. This paper proposes a no v el nighttime underw ater sea turtle detector with vision approaches, named NiteT ron. Based on an ef cient modication of Y OLO11-nano, called Y OLO11-pico. The model achie v es competiti v e performance compared to other rec ent Y OLO architectures, which utilize lightweight con v olu- tional modules to impro v e detection accurac y in lo w-light en vironments. The netw ork introduces tw o main modules to enhance the backbone. Strong feature modulation (SFM) captures a lar ger recepti v e eld using a big k ernel and applies a g ate operation to select rele v ant cells. It also fuses information and strengthens feature relationships. Furthermore, sparse separated con v olution (S2C) ef ciently e xtracts characteristic fea tures of turtles by combining partial and depthwise con v olutions. It also inte grates enhancing channel (EC) modules, which use broadcast multiplication to emphasize rele v ant feature channels. The main contrib utions of this w ork are summarized as follo ws: - A no v el sea turtle detection method is proposed for nighttime conditions, utilizing an ef cient deep learning approach that optimizes both detection performance and inference speed. - A ne w ef cient v ariant of the Y OLO11 model is introduced, utilizing a series of lightweight operations and a reduced layer structure to ef fecti v ely e xtract discriminati v e features while minimizing trainable parameters and computational comple xity . - A ne w ef cient feature e xtraction block is proposed, emplo ying lightweight con v olutional layers. This mod- ule partitions the input fe ature map and inte grates a channel enhancement mechanism at each layer , ef fec- ti v ely reducing the number of trainable parameters and computational cost. - A no v el enhancement modulation module is of fered, implementing sequential attention blocks that include channel attention and g ated operations. This module enhances the feature e xtraction capability by capturing essential conte xtual information and reinforcing the relationships between spatial and global features. Its plug-and-play design encourages a rob ust and adaptable attention module suitable for inte gration into deep learning architecture. - A comprehensi v e model analysis is conducted to e v aluate the ef fecti v eness and ef cienc y of the proposed lightweight and enhancement module. The study assesses both performance and computational o v erhead introduced by the additional blocks. Furthermore, the proposed detector is e v aluated on nighttime sea turtle TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 5, October 2026: 1561–1580 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1563 imagery from Bunak en and demonstrates competiti v e performance in mAP compared to other lightweight Y OLO-based detectors. T o v alidate the ef cienc y of the proposed model, additional metrics are reported, including the number of parameters, GFLOPs, and runtime speed on dif ferent central processing unit (CPU) hardw are platforms. 2. RELA TED W ORKS Underw ater object detection using CNN models has been e xtensi v ely applied, with a primary fo- cus on impro ving detection accurac y . Pre vious w orks, such as attention-free h ybrid netw ork (AFHNet) [11], Y OLOv8-UC [12], HLAST -A CNet [13], and enhanced Y OLOv5 [14]. The y ha v e demonstrated that CNN- based architectures can achie v e real-time underw ater object detection with impro v ed performance. Li et al. [15], study proposed an underw ater detector based on the real-time detection transformer (R T -DETR) archi- tecture, called Underw ater -DETR (UW -DETR), which outperformed se v eral R T -DETR v ariants. Ho we v er , it requires high computational resources and a lar ge number of parameters. Additionally , Y OLOv7-PE [16] im- pro v es the balance between accurac y and ef cienc y . It is b uilt upon the Y OLOv7 architecture, incorporating the con v olutional block attention module (CB AM) module to enhance detection accurac y and the cross stage partial spatial p yramid pooling-f ast (CSPSPPF) module to reduce computational cost. Ne v ertheless, its perfor - mance has not yet been e v aluated under lo w-light conditions. Furthermore, Zhang et al. [17] introduced an ef cient underw ater object detector based on Y OLOv8, with fe wer parameters and lo wer GFLOPs compared to Y OLOv8-Lar ge. Ho we v er , the model e xhibits limitations in handling o v erlapping objects, which may result in missed detections of underw ater objects. Moreo v er , deep learning methods ha v e signicantly adv anced marine animal detection for ecosyst em monitoring. Y OLOv7 w as emplo yed as an object detector within a weight predict ion system to assist in turtle sorting [18]. Similarly , Zakry et al. [19] applied deep learning techniques to sea turtle detection, supporting conserv ation ef forts through more ef cient and real-time monitoring. Ne v ertheless, grayscale images used in model training pose challenges for deplo yment in real-time applications using red, green, and blue (RGB) cameras. The study in Baek et al. [20] conducts a comparati v e analysis using Y OLOv5-se g for sea turtle classication with instance se gmentation, highlighting its potential for precise object delineation, though it does not focus on model optimization or deplo yment ef cienc y . Another study , Liu et al. [21] proposes global attention mechanism and transfer learning (PLGA T), an underw ater Plectr opomus leopar dus detector based on a modied Y OLOv8. It achie v es 0.958 mAP on t he Plectr opomus leopar dus r eco gnition dataset (PLRD) dataset and 0.974 on the Brackish dataset. Ho we v er , it requires 7.4 million parameters, making it unsuitable for edge de vices. As k e y contrib utors to marine ecosystems, sea turtles ha v e become a focus of research in popula tion monitoring and conserv ation. Adam et al. [22] introduced a long-span dataset for sea turtle re-identication based on stable e xternal features, enabling continuous indi vidual tracking o v er time. Ho we v er , the recognition system is not designed for video-based or real-time deplo yment. Additionally , pre vious w o r k [2] proposed a modied Y OLOv8-nano netw ork for sea turtle detection, introducing the e xci ted group attention (EGA) mod- ule to generate multi-sca le representations across v arious recepti v e elds. The modul e emplo ys channel shuf e operations to enhance feature blending. Using this approach, the proposed model achie v ed a mAP50:95 score of 0.616, representing a 0.049 impro v ement o v er the original Y OLOv8-nano netw ork. This model also reduces the number of channels compared to the original Y OLOv8-nano, making it more lightweight and ef cient. An- other recent study , Putro et al. [3] presented an ef cient sea turtle detector by limiting the maximum number of channels in the netw ork. This modication to Y OLOv8 reduces the parameter count by 1.1 million. Addi- tionally , the authors propose dual-spatial enhancement dan gradual recepti v e enhancement modules to impro v e detection accurac y , achie ving a mAP50:95 score of 0.604. T o enable continuous sea turtle monitoring, the proposed method must handle lo w-light conditions at night, which reduce distinguishing features. Se v eral studies ha v e e xplored nighttime object detection [23]- [25]. Zhao et al. [26] proposed a Y OLOv5-based approach for nighttime pedestrian detection usi ng infrared cameras to capture lo w-light en vironments, achie ving a mAP score of 93.62. Ne v ertheless, the method is limited to infrared imagery . Further e v aluation on RGB data is required to ensure more e xible and practical deplo yment. Additionally , Xu et al. [27] introduces a nighttime v ehicle detection approach to support the de v elopment of intelligent tra nsportation systems. The method enhances f aster re gion-based con v olutional neural netw ork (R-CNN) by inte grating deformable con v olutional netw orks, achie ving a mAP score of 83.3. NiteT r on: a no vel nighttime b unak en sea turtle species detection (Muhamad Dwisnanto Putr o) Evaluation Warning : The document was created with Spire.PDF for Python.
1564 ❒ ISSN: 1693-6930 Ho we v er , the e v aluation is limited to a s mall-scale, single-s cene dataset, which does not reect the comple xity and dark er conditions often encountered in real-w orld scenarios. Existing approaches e xhibit limitations under lo w-light conditions, as reduced illumination de grades feature visibility and ne g ati v ely impacts detection performance. Notably , Putro et al. [3] e v aluates the model e xclusi v ely under daylight conditions, despite util izing the same baseline dataset as this study . Therefore, this study proposes a nighttime underw at er sea turtle detection model, named NiteT ron, b uilt on a modied Y OLO11-nano architecture referred to as Y OLO11-pico. The model is optimized for reliable performance in lo w-light conditions using standard RGB camera s, maki ng it suitable for deplo yment on a wide range of de vices. Additionally , the reduced number of channels mak es the architecture both lightweight and ef cient. Furthermore, tw o no v el modules are incorporated to impro v e detection accurac y through lightweight enhance- ments while maintaining computational ef cienc y . The proposed method incorporates an ef cient architecture that achie v es competiti v e performance when compared to recent Y OLO v ariants. In addition, this study focuses on the detection and recognition of sea turtles in Bunak en Marine P ark, specically tar geting three species, namely green, oli v e ridle y , and ha wksbill. 3. PR OPOSED ARCHITECTURE This w ork proposes tw o no v el modules in the main e xtractor , namely SFM and S2C, as part of the lightweight NiteT ron architecture, resulting in an ef cient nighttime underw ater sea turtle detector . The S2C module is desi gned to e xtract distinguishing features through lightweight operations, while the SFM module captures a lar ge recepti v e eld. These modules enhance the netw ork’ s performance while maintaining ef - cienc y . This design enables the model to achie v e real-time inference speed on promising de vices. 3.1. Backbone The backbone is a k e y component of Y OLO11 architecture, res ponsible for e xtracting features from input images using con v olutional layers and attention modules. As illustrated in Figure 1, it emplo ys 3 × 3 con v olutional blocks that function both as feature e xtractors and do wnsampling units when it uses a stride of 2 to reduce the spatial di mensions of the feature maps. Each con v oluti onal layer is accompanied by batch normalization to stabilize data distrib ution across batches and channels, follo wed by the sigm o i d linear unit (SiLU) acti v ation function, which introduces non-linearity while retaining small ne g ati v e v alues. Other than that, the C3K2 module plays a central role in feature e xtraction by pro viding a f aster implementation of the cross stage partial (CSP) structure. It inte grates the C3K module, which consists of three con v olutional layers and includes an S2C unit to enhance feature quality . In addition, C3K is a module based on the bottleneck architecture, designed to support customizable con v olutional k ernel sizes. It splits the input feature maps along the channel dimension to enable ef cient information e xtraction, using a 1 × 1 con v olution to dene the channel partitions. One-half of the channels passes through a bottleneck that replaces the standard con v olution with the proposed S2C module. The output from the S2C module is fused using concatenate operation with the identity branch, which contains the remaining channels, using a concatenation operation. A nal 1 × 1 con v olution is applied to mix the concatenated feature maps and to capture channel relationships. Furthermore, the spat ial p yramid pooling f a ster (SPPF) module performs feature aggre g ation. It in- cludes tw o 1 × 1 con v olutions, positioned at the be ginning and end of the process, along with three sequential a v erage pooling operations using a 5 × 5 windo w in the middle of the process. Ho we v er , the second 1 × 1 con- v olution is remo v ed. The rst 1 × 1 con v olution is emplo yed to set the number of channels before aggre g ation. Using multiple 5 × 5 pooling operations in sequence allo ws the model t o capture increasingly lar ger recepti v e elds. F or instance, tw o 5 × 5 layers approximate a 9 × 9 recepti v e eld, while three layers approximate a 13 × 13 eld. The resulting feature m aps from each scale are combined to form a rich multi-scale representation using a concatenation operation. The SFM module is placed at the end of the backbone to enhance channel- wise information. Subsequently , a 1 × 1 con v olution is applied at the end to mix the channel-wise information deri v ed from the SFM output, further emphasizing the most rele v ant features for detection. In order t o impro v e ef cienc y , reducing the number of channels in se v eral layers is essential for a v oiding redundant feature maps and lo wering computational cost. This reduction helps the map concentrate on meaningful object information. Among all Y OLO v ersions, the nano v ariant is the smallest, as it utilizes fe wer parameters and GFLOPs for operation. Ho we v er , it still presents opportunities for further optimization, particularly for f aster inference on CPU-based de vices. This w ork proposes a ne w Y OLO architecture, referred to as Y OLO-pico, which adopts a more compact design by using fe wer param eters and lo wer GFLOPs than TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 5, October 2026: 1561–1580 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1565 Y OLO-nano. This compactness results from a deliberat e reduction in the number of output channels across layers. The complete channel conguration is sho wn in Figure 1. Th e results demonstrate that increasing the number of channels does not necessarily lead to better accurac y . Instead, selecti v ely retaining only essential channels yields a more ef cient trade-of f between accurac y and computational performance. Figure 1. The proposed NiteT ron architecture is based on Y OLO11-pico and inte grates tw o enhancement modules into the backbone. It emphasizes ef cient feature e xtraction and channel renement to impro v e detection performance 3.2. S2C A backbone relies on con v olutional operations to e xtract object information from input images, cap- turing local spatial features from the resulting feature maps. The original C3K2 module incorporates C3K (three con v olutions with adapti v e k ernels), which includes a bottle n e ck structure that applies a 3 × 3 k ernel twice, ef fecti v ely approximating the recepti v e eld of a 5 × 5 k ernel. Ho we v er , identifying which feature maps contain meaningful object information requires global conte xt across channel representations, which typically demands a lar ge number of parameters to model long-range dependencies. In this conte xt, S2C aims to ef - ciently capture local spatial features by applying tw o types of con v olution in separate processes with global representation. This mechanism relates spatial-le v el data without increasing the use of parameters and compu- tational resources. T o highlight essential feature maps, the captured global representat ion is used as a scaling f actor for the corresponding feature maps, allo wing them to be adjusted based on their global v alues. It is implemented at the medium and high-le v el feature stage to ensure that the most rele v ant feat u r es are enhanced while suppressing less information. As sho wn in Figure 2, the S2C module inte grates standard con v olution and depthwise con v olution, which applies lters across all input channels and processes each half of the channel independently . The operation can be formulated as (1): S 2 C ( X ) = C v ( E C ( C oncat [ C v s 1 ( X ) , C v s 2 ( X )])) (1) The input feature maps X are split into tw o halv es. Using only standard con v olution t o e xtract features w ould increase the number of parameters and computational cost. T o address this, S 2 C applies dif ferent NiteT r on: a no vel nighttime b unak en sea turtle species detection (Muhamad Dwisnanto Putr o) Evaluation Warning : The document was created with Spire.PDF for Python.
1566 ❒ ISSN: 1693-6930 con v olution types to each half of the channels. A standard con v olution ( C v s 1 ( X ) ) is applied to one half to preserv e inter -channel relationships, while a depthwise con v olution ( C v s 2 ( X ) ) is utilized to the other half to enhance computational ef cienc y . The o v erall con v olution process is illustrated as (2): C v s 1 = C v 3 × 3 ( X s 1 ) , C v s 2 = D w 3 × 3 ( X s 2 ) (2) T o preserv e local spatial relationships, a 3 × 3 k ernel is emplo yed, striking a balance between ef - cienc y and neighborhood feature representation. Half of the channels are processed using C v 3 × 3 , which con- sists of a 2D standard con v olution with 3 × 3 k ernel size follo wed by batch normalization to stabilize feature distrib ution, and a SiLU acti v ation function to introduce non-linearity . While standard con v olution captures inter -channel dependencies, the remaining half of the channels are processed using depthwise con v olution with the same 3 × 3 k ernel, also follo wed by batch normalization and SiLU acti v ation. The tw o branches, X s 1 and X s 2 , represent the split feature maps obtained from the output of the con v olution operation C v , which utilizes a 1 × 1 k ernel along with batch normalization and SiLU acti v ation, denoted as (3). [ X s 1 , X s 2 ] = C v (3) This splitting design impro v es computational ef cienc y while preserving accurate feature e xtraction. Each branch follo ws a distinct processing strate gy . Furthermore, the feature maps generated from dif ferent con v olution branches are concatenated along the channel di mension to form a unied tensor , allo wing features from dif ferent processing branches to be jointly processed. This fusion enhances the ability of feature e xtraction by capturing relationships between feature maps. T o strengthen the interaction between outputs from standard and depthwise con v olutions, a con v olutional layer ( C v ) is applied. Figure 2. S2C, this block applies separate con v olutions to obtain dif ferent feature representations, which are enhanced with attention channels This layer consists of a 1 × 1 con v olution, follo wed by batch normalization and a SiLU acti v ation function. It focuses on EC-wise relationships, making the operation ef cient for this task. Additionally , this TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 5, October 2026: 1561–1580 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1567 step restores the output to t he intended number of channels, ensuring compatibility wit h the ne xt stages of the netw ork. In addition, impro ving the recognition of informati v e features requires identifying which channels carry the most signicant information. Channel attention mechanisms commonly le v erage global pooling to summarize each feature map into a representati v e v alue, which acts as a scaling f actor for enhancing rele v ant channels. In the S 2 C module, global a v erage pooling ( GAP ) is used to perform this operation, preserving the original distrib ution of feature v alues. This process is dened as (4). E C ( X ) = GAP ( X ) ∗ X (4) The scaling operation impro v es feature map quality by emphasizing channels that carry more rel e v ant object information. GAP e xtracts representati v e v alues from each channel, which serv e as scaling f actors for the input feature maps. This process represents EC that emplo ys broadcasting multiplication via scaling v alues uniformly across the corresponding channels. 3.3. SFM Attention mechanisms are ef fecti v e techniques for enhancing model accurac y by enabling neural net - w orks to focus on the m ost informati v e parts of an input. These modules rene feature representations by emphasizing specic feature maps that are most rele v ant to the object of interest. Ho we v er , man y e xisting attention mechanisms operate within a single dom ain, either channel-wise or spatially , which limits their ef fec- ti v eness in comple x or lo w-visibility scenarios. Channel attention modules capture global conte xt by modeling the relationships across channels. While this helps the model prioritize semantically rich feature maps, it often misses spatial detail that is crucial for detecting small or ne-grained object features. On the other hand, spatial attention modules focus on identifying important re gions within the feature map b ut tend to be computationally intensi v e, especially when capturing interactions across channels or operating on high-resolution data. T o impro v e detection accurac y in challenging scenarios, it is essential to incorporate both channel and spatial attention mechanisms. At the same time, maintaining lo w computational cost is critical to preserving o v erall model ef cienc y . In this w ork, we introduce a lightweight dual-attention module called SFM, which adopts the star technique [28] to enhance fea ture representation without incurring signicant o v erhead. SFM inte grates both channel and spatial attention into a unied, ef cient structure. It combines a squeeze-and- e xcitation (SE) layer , which captures inter -channel dependencies, with a lar ge conte xt area (LCA) block that models spatial information using lar ge k ernel depthwise con v olutions. The outputs from both branches are fused using element-wi se multiplication, enabling the netw ork to emphasi ze features based on both their spatial location and channel importance. By combining these complementary forms of attention, SFM pro vides a more rob ust and discriminati v e feature enhancement mechanism. This mechanism is particularly ef fecti v e in visually de graded en vironments, such as nighttime underw ater scenes, where object features are often subtle, and spatially scattered. As sho wn in Figure 3, the SFM module comprises se v eral components, be ginning with a SE [29] block that assigns higher importance to the most informati v e feature maps by generating channel-wise prob- ability weights. The output from the SE block is then normalized using layer normalization (LN), which stabilizes the feature distrib ution across batches and reduces the impact of redundant and outl ier v alues. The normalized features are then forw arded into tw o parallel branches, both operating on the same input to e x- tract complementary information for further renement. T o capture a wider recepti v e eld while maintaining computational ef c ienc y , one branch applies a depthwise con v olution with a 7 × 7 k ernel. On the other hand, the second branch applies a 1 × 1 con v olution to model inter -channel relationships and introduce v ariation in feature transform ation. This dual-branch structure enables SFM to ef fecti v ely combine spatial and channel information, thereby enhancing feature representation. The outputs from both the LCA and 1 × 1 con v olution branches are then combined using element-wise multiplication to mer ge spatial and channel-a w are features. This fused representation is further rened through the feed forw ard with fusion (F3) module, which enhances feature mixing and reinforces channel relationships. The result is passed through a nal 1 × 1 con v olution to produce the output of the SFM module. The o v erall process can be dened as (5): S F M ( X ) = C v ( F 3( LC A ( S E N ( X )) C onv 1 × 1 ( S E N ( X )))) (5) Squeeze and e xcitation module with normalization (SEN) is a widely adopted attention mechani sm kno wn for its ability to impro v e model performance by adapti v ely recalibrating channel-wise feature informa- tion. It utilizes a v erage pooling to squeeze feature maps dimension into 1 × 1 × C and uses linear operation to NiteT r on: a no vel nighttime b unak en sea turtle species detection (Muhamad Dwisnanto Putr o) Evaluation Warning : The document was created with Spire.PDF for Python.
1568 ❒ ISSN: 1693-6930 e xtract information based on channels. This attention module is applied at the be ginning to assess the rele v ance of each feature map and assign higher weights to those containing more meaningful object information. This process can be illustrated as (6). S E ( X ) = ( W 2 ( R eLU ( W 1 ( GAP ( X ))))) ∗ X (6) Figure 3. SFM, this module enhances feature representations by combining spatial and channel attention, allo wing the netw ork to focus on informati v e features under lo w-light conditions Channel attention f o c uses on identifying which channels contain the most rele v ant information. GAP is an ef fecti v e method for obtaining a representati v e v alue for each channel by computing the a v erage of all spatial elements within a feature map, resulting in a single scalar v alue per channel. This operation captures global conte xt and reects the importance of each channel relati v e to the input. T o further enhance feature representation, a tw o-layer fully connected netw ork is applied after the pooling operation. The rst linear layer reduces the dimensionality using a reduction ratio of 16 to minimize computational cost. A rectied linear unit (ReLU) acti v ation is applied to suppress non-informa ti v e or ne g ati v e v alues. The second linear layer restores the original channel dimension, follo wed by a sigmoid acti v ation that generates attention weights. These weights guide the model to focus on channels that carry more informati v e features. This process enables the model to focus on the most informati v e features, making the subsequent feature e xtraction more ef cient and ef fecti v e in capturing critical object details. T o preserv e feature v ariation and ensure stable data distrib ution between each batches, the model incorporates LN immediately after the SE module. This operation can be described as (7). S E N ( X ) = LN ( S E ( X )) (7) Similar to other channel attention mechanisms, the SE module reduces quality of spatial information, potentially discarding ne-grained details across e xtended spat ial re gions. T o o v ercome this limitation, the de- sign incorporates a LCA module, which is responsible for capturing and reinforcing spatial i n f ormation through a dual-branch structure. The rst branch applies a pointwise con v olution to introduce feature v ariation and en- rich the e xtracted representations. In parallel, the second branch uses a lightweight sequential con v olution to capture a broader recepti v e eld and strengthen spatial relationships. This process is formulated as (8): LC A ( X ) = D W 7 × 7 ( GeLU ( C onv 1 × 1 ( X ))) (8) TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 5, October 2026: 1561–1580 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1569 The LCA module aims to capture rich spatia l conte xt while minimizing computational cost. It be g i ns with a 1 × 1 standard con v olution ( C onv 1 × 1 ), follo wed by Gaussian error linear unit (GELU) acti v ation. This step enhances inter -channel interactions and prepares the feature maps for more ef fecti v e spatial processing. This non-linear acti v ation is critical for rened feature selection. Unlik e ReLU, which discards all ne g ati v e v alues, GELU applies a smooth g ating mechanism that allo ws small ne g ati v e v al ues to pass through. This mechanism helps recei v e small ne g ati v e v alues that ha v e potentially meaningful informa tion. In order to enrich spatial understanding, a 7 × 7 depthwise con v olution ( D W 7 × 7 ) is appli ed. The lar ge k ernel size allo ws the netw ork to capture broader spatial dependencies, while the depthwise design helps control parameter gro wth and maintain computational ef cienc y . This combination enables the LCA module to model detailed local structures and long-range conte xt ef ciently . Furthermore, the outputs from the C onv 1 × 1 and LCA branches are then fused using element-wise multiplication, reinforcing the most informati v e features and enhancing the o v erall feature representation. Additionall y , to enhance feature interactions, the Feed F orw ard with Fusion ( F 3 ) module is emplo yed. This module renes the fused representation by emphasizing inform ati v e features a nd reducing channel redundanc y . It utilizes a single spatial con v olution to st rengthen inter -channel relationships, follo wed by an attention module that highlights channels containing the most rele v ant information based on the enhanced features. This sequence of operations is represented as (9): F 3( X ) = C onv ( R eLU ( C oncat [ S E ( C onv ( X )) , C onv ( X )])) (9) The F3 module be gins with a 1 × 1 con v olution ( C onv ) to establish relationships between the input features, enhancing the representation of rele v ant informati on. It is follo wed by a S E block that recalibrates channel-wise importance and amplies the channels that cont ain the most object-rele v ant information. The output of the SE block is concatenated wit h the ( C onv ) feature using a concatenation operation ([·]), ef fecti v ely doubling the channel dimension and enabling richer feature interaction. The combined tensor serv es as a unied representation for subsequent processing. T o eliminate this irrele v ant information, a ReLU function is applied, which remo v es all ne g ati v e v alues. This operation retains only positi v e responses, allo wing the netw ork to focus on features that are more lik ely to contrib ute meaningful information. Finally , another 1 × 1 con v olution ( C onv ) reduces the feature dimension back to its original size, producing a compact and enriched output that preserv es both spatial and semantic i nformation. This laye red processing within the F3 module enabl es the netw ork to adapti v ely fuse and rene features after a ttention, ensuring t hat the output of the SFM block is both discriminati v e and lightweight. 3.4. Neck The feature map o w within a netw ork is fundamental to ef fecti v e learning and accurate pre diction, particularly in object detection tasks. As illustrated in Figure 1, Y OLO11-pico architecture maintains the core structure of i ts predecessor , Y OLO11-nano, and adopts the path aggre g ation netw ork (P AN) [30] as its neck component. The P AN structure plays a crucial role in fusing multi-scale feature maps generated by the backbone, allo wing the model to inte grate both lo w-le v el spatial details and high-le v el semantic conte xt. The neck emplo ys a concatenation approach to combine feature maps from dif ferent stages. This operation enables the preserv ation of detailed information from lo w-le v el layers, which is essential for detecting small objects. By f acilitating a rich information o w between features of v arious scales, the P AN structure enhances the di v ersity of the feature representations passed to the detection head. This mechanism enables better localization and classication performance while maintaining a compact and ef cient architecture. 3.5. Head Y OLO11-pico retains the detection head structure and loss functions from the original Y OLO11- nano. The head features tw o parallel branches, separating bounding box re gression and object classication tasks. This decoupled design allo ws the model to optimize each task independently , while maintaining ef - cienc y and s uitability for real-time and edge deplo yment. The detection head combines standard and depthwise con v olutional layers. In the bounding box re gression branch, it appli es tw o 3×3 s tandard con v olutions, each follo wed by batch normalization and a SiLU acti v ation. A nal 1×1 con v olutional layer generates the bound- ing box predictions. It produces tensor outputs at three spatial resolutions, where each scale corresponds to a dif ferent l e v el of fea ture representation, from lo w-le v el to high-le v el features. Each output tensor has a channel dimension equal to four times the v alue of re gmax. The bounding box re gression predicts four v alues repre- senting the diagonal coordinates of the bounding box, specically xmin, ymin, xmax, and ymax. The re gmax parameter denes the maximum range for re gression and is set to 16 by def ault. NiteT r on: a no vel nighttime b unak en sea turtle species detection (Muhamad Dwisnanto Putr o) Evaluation Warning : The document was created with Spire.PDF for Python.
1570 ❒ ISSN: 1693-6930 The classication branch is responsible for predicting object classes and uses a deeper con v olutional structure compared to the re gression branch. It applies tw o sequential operations, each consisting of a 3×3 depthwise con v olution follo wed by batch normalization and SiLU acti v ation, and then a 1×1 standard con v o- lution that is also follo wed by batch normalization and SiLU acti v ation. Additionally , a nal 1×1 con v olution layer is placed at the end to produce the classication output. Similar to the bounding box branch, it gen- erates outputs at three grid sizes. Ho we v er , the num b e r of output channels in the classi cation branch is set equal to the number of classes to represent class probabilities ef ciently . This conguration captures spatial and channel-le v el features ef fecti v ely while k eeping the computational load lo w . The training stage utilizes a composite loss function that combines complete intersection o v er union (CIoU) for precise localization [31], distrib ution focal loss (DFL) [32] for distrib uted bounding box prediction, and binary cross-entrop y (BCE) [33] for classication supervision. 3.6. Dataset This study utilizes a dataset from pre vious w ork [3], collected in Bunak en Marine P ark for the de v elop- ment of a sea turtle detection model. The dataset includes annotated images of three sea turtle species , namely green, ha wksbill, and oli v e ridle y , captured under v arious underw ater conditions. The dataset is augmented using a handcrafted process with image editing softw are to simulate nighttime scenes by reducing brightness to -124, s etting contrast to zero, adjusting e xposure to -4.19, and applying a g amma of one. It als o incorporates localized lighting ef fects that mimic articial illuminat ion, as sho wn in Figure 4. These augmentations help replicate the lo w-light conditions commonly encountered during nighttime monitoring. The dataset used in this study includes real (non-augmented) nighttime images captured directly from Bunak en w aters under nat- ural lo w-light conditions. These images reect real-w orld challenges such as limited illumination, underw ater noise, and visibility de gradation. Although augmentation techniques were applied to enhance rob ustness, the model is primarily trained on authentic nighttime data to ensure realistic feature representation. The baseline dataset w as collected from multiple di v e sites within Bunak en National P ark. These sites were selected based on the re gional sea turtle distrib ution map, which identies areas with high turtle acti vity and biodi v ersity . These locations represent ecologically di v erse underw ater en vironments, v arying in reef structures, w ater clarity , and natural lighting conditions. Subsequently , the dataset is annotated manually using a bounding-box tool, generating Y OLO formatted labels that i n c lude the cl ass label and normalized coordinates (x, y , w , h). The complete dataset comprises 48,244 annotated images, totaling 5.4 GB, with v arying r esolutions of 3840 × 2160, 1920 × 1080, and 1080 × 720. The dataset is di vided into three subsets to support model de v elopment. Out of the total images, 30,934 were allocated for training, 7,698 for testing, and 9,612 for v alidation. The dataset comprises 28,053 images of Chelonia mydas (green turtle), 11,382 of Er etmoc helys imbricata (ha wksbill turtle), and 8,809 of Lepidoc helys olivacea (oli v e ridle y turtle). It is locally collected and class-imbalanced, and no data augmentation is applied to increase the pattern recognition challenge. The training set enables the model to learn and update its parameters, while the v alidation set guides h yperparameter tuning and performance tracking during training. The test set pro vides a nal benchmark to e v aluate ho w well the model generalizes to unseen data. This structured split supports a balanced and reliable learning process. Figure 4. V isualization of underw ater dataset augmentation. Handcrafted augmentations simulate nighttime conditions by reducing brightness, adjusting contrast, and adding lighting ef fects 3.7. T raining conguration In order to ensure a f air comparison, the training process does not emplo y an y transfer learning meth- ods. The NiteT ron model is trained for 150 epochs, which is identied as the optimal conguration based on empirical analysis, as training be yond this point yields di minishing performance g ains and increases the risk of v anishing gradients. Furthermore, both tr aining and e v aluation used an input resolution of 640 × 640 TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 5, October 2026: 1561–1580 Evaluation Warning : The document was created with Spire.PDF for Python.