IAES Inter national J our nal of Robotics and A utomation (IJRA) V ol. 15, No. 3, September 2026, pp. 544 ∼ 552 ISSN: 2722-2586, DOI: 10.11591/ijra.v15i3.pp544-552 ❒ 544 Spatial-channel r econstruction f or efcient multiscale attention in r obotic object detection Mohammed Maiza 1 , Chahira Cherif 2 , Samira Chouraqui 3 , Abdelmalik T aleb-Ahmed 4 1 LITIO Laboratory , F aculty of Exact and Applied Sciences, Uni v ersity of Oran 1 Ahmed Ben Bella, Oran, Algeria 2 RIIR Laboratory , F aculty of Medicine, Uni v ersity of Oran 1 Ahmed Ben Bella, Oran, Algeria 3 F aculty of Mathematics and Computer Science, Uni v ersity of Sciences and T echnology of Oran, Oran, Algeria 4 IEMN, Polytechnic Uni v ersity of Hauts-de-France, Uni v ersity of Lille, V alenciennes, France Article Inf o Article history: Recei v ed Feb 17, 2026 Re vised Apr 24, 2026 Accepted May 22, 2026 K eyw ords: Autonomous robots Lightweight attention Object detection Real-time perception Y OLO ABSTRA CT Real-time object detection is a core capability for autonomous robots, un- manned aerial v ehicles (U A Vs), and self-dri ving systems operating in resource- constrained en vironments. This paper presents spatial-channel enhanced mul- tiscale attention (SCEMA), a no v el l ightweight attention module designed to enhance robotic perception while minimi zing computational o v erhead for em- bedded deplo yment. SCEMA emplo ys a parallel dual-branch architecture that syner gistically combines spatial-channel reconstruction with multiscale atten- tion mechanisms. When inte grated into the Y OLOv8n frame w ork (3.01M base- line parameters), the proposed Y OLO-SCEMA model achie v es signicant per - formance g ains across multiple challenging benchmarks rele v ant to robotics au- tomation. Experiments on an NVIDIA R TX 4080 GPU demonstrate that on the ExDark datase t, Y OLO-SCEMA impro v es mAP@50 by 7.37% o v er the baseline (69.07% to 76.44%) while reducing parameters by 36.88% (3.01M to 1.90M) and computational cost by 8.64% (8.1 to 7.4 GFLOPs). Consistent im- pro v ements are also observ ed on V isDrone2019 (+3.24% mAP@50) and FYP (+1.50% mAP@50) datasets. Comparati v e analysis demonstrates that Y OLO- SCEMA achie v es superior accurac y-ef cienc y trade-of fs, making it particularly suitable for deplo yment in lo w-light conditions, dense scenes, and comple x structural en vironments for autonomous na vig ation, robotic surv eillance, and industrial automation applications. This is an open access article under the CC BY -SA license . Corresponding A uthor: Mohammed Maiza LITIO Laboratory , F aculty of Exact and Applied Sciences, Uni v ersity of Oran 1 Ahmed Ben Bella Oran, Algeria Email: maiza.mohammed@uni v-oran1.dz 1. INTR ODUCTION Object detection constitutes a fundamental perception capability for autonomous robotic s ystems, including ground robots, unmanned aerial v ehicles (U A Vs), and self-dri ving v ehicles. These systems must reli- ably detect and localize objects in real-time while operating under se v ere computational and ener gy constraints inherent to embedded robotic platforms [1], [2]. Con v olutional neural netw orks (CNNs) ha v e become the dom- inant paradigm for visual perception in robotics; ho we v er , modern detectors achi e v e high accurac y at the cost of increased depth and width, signicantly raising computational and memory requirements that limit deplo yment on resource-constrained robotic hardw are [3]. In robotic applications such as autonomous na vig ation, surv eil- lance, and industrial automation, perception systems f ace three critical challenges: i) lo w-light conditions (e.g., J ournal homepage: http://ijr a.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 545 nighttime autonomous dri ving), ii) dense object distrib utions with occlusions (e.g., U A V -based cro wd mon- itoring), and iii) comple x structural scenes (e.g., manuf acturing en vironments). As netw ork depth increases, feature maps e xhibit substantial spatial and channel redundanc y [4], de grading computational ef cienc y . While attention mechanisms ha v e been introduced to emphasize informati v e features, man y e xisting approaches incur non-ne gligible computational o v erhead or focus on either spatial or channel dependencies in isolation. Con- sequently , designing lightweight attention modules that jointly e xploit spatial and channel interactions while preserving multiscale conte xtual information remains a critical challenge for real-time robotic perception [5], [6]. Recent Y OLO-based detectors emphasize ef cienc y and real-ti me performance, yet their lightweight v ari- ants often suf fer from de graded ac curac y in challenging robotic en vironments. T o address this g ap, this paper proposes SCEMA, a no v el lightweight attention module that enhances robotic perception without introducing e xcessi v e computational cost. This w ork adv ances robotics and automation by enabling lightweight, ef cient, and rob ust per ception for autonomous systems. The paper is or g anized as follo ws. Section 2 re vie ws related w ork on object detection architectures and attention mechanisms in r ob ot ics conte xts. Section 3 presents the SCEMA methodology , in- cluding inte gration into robotic vision pipelines. Section 4 details e xperimental results across three benchmark datasets, with analysis framed in terms of robotic metrics. Finally , section 5 concludes with implications for autonomous robotics and future directions. 2. RELA TED W ORK Attention mechanisms enhance feature discrimination by selecti v ely emphasizing informati v e aspects of input data, which is particularly v aluable for robotic perception in cluttered or lo w-visi bility en vironments. Channel attention methods, pioneered by SENet [7], emplo y global a v erage pooling follo wed by fully con- nected layers to recalibrate feature responses. Spatial attention approaches compute importance maps across spatial dimensions, highlighting re gions with high semantic rele v ance. Hybrid attention modules, such as CB AM [8] and ECA [9], combine spatial and channel cues sequentially or jointly . Ho we v er , man y attention modules introduce computational o v erhead through parameter -intensi v e operations, hindering deplo yment on robotic hardw are with limited processing capabilities. Furthermore, e xisting mechanisms frequently f ail to ef- ciently model multiscale spatial dependencies, limiting their ef fecti v eness for robotic applications in v olving objects at v arying distances and aspect ratios. The demand for real-time object detection in robotics has dri v en research into lightweight detectors that balance computational ef cienc y and detection accurac y . The Y OLO f amily has emer ged as a dominant paradigm for real-time robotic perception due to its single-stage design and ef cient feature e xtraction backbone [10]. Recent v ariants, incl uding Y OLOv8 [11], Y OLOv9 [12], Y OLOv10 [13], Y OLOv11 [14], and Y OLOv12 [15], ha v e progres si v ely impro v ed the speed-accurac y trade-of f. Ho we v er , aggressi v e model compression can signicantly de grade performance in challenging scenarios characteristic of robotic applications: lo w lighting, dense object distrib utions, occlusions, and signicant scale v ariations. De- v eloping lightweight detectors that maintain rob ust performance across di v erse en vironments remains critical for autonomous na vig ation, U A V surv eillance, and industrial automation. Multi scale feature representation is essential for robotic perception, as real-w orld scenes contain objects with substantial size v ariations. Feature p yramid-based architectures [16] construct hierarchical representations through top-do wn fusion or bidirec- tional information o w . Ho we v er , these me thods often rely on computationally e xpensi v e fusion strate gies that increase inference latenc y and memory consumption, challenging real-time robotic operation. The in- te gration of attention mechanis ms with multiscale feature p yramids remains undere xplored, limiting potential syner gistic benets. Ef cient multiscale attention mechanisms that operate across dif ferent feature scales while maintaining real-time performance represent a signicant open research problem for robotics [17], particularly for applications requiring long-range dependenc y modeling and conte xtual inte gration from dif ferent recepti v e elds. 3. METHOD The o v erall architecture of SCEMA is illustrated in Figure 1. SCEMA adopts a parallel proce ssing strate gy consisting of tw o complementary branches. In the l eft branch, spatial-rened features X w are e xtracted using the spatial renement unit (SR U) and spatial attention, follo wed by channel renement through the channel renement unit (CR U) to produce channel-rened features Y . The right branch focuses on learning ef fecti v e channel representations by grouping features without Spatial-c hannel r econstruction for ef cient multiscale attention in r obotic object (Mohammed Maiza) Evaluation Warning : The document was created with Spire.PDF for Python.
546 ❒ ISSN: 2722-2586 reducing channel dimensions, enhancing pix el-le v el attention and rob ustness to multi-scale v ariations. The architecture of the Y OLO-SCEMA model is sho wn in Figure 2. Figure 1. Architecture of the proposed SCEMA module Figure 2. Ov erall architecture of the Y OLO-SCEMA model SCEMA is inte grated into the Y OLOv8 backbone, forming a perception module suitable for robotic autonomy stacks. W ithin the con v olutional block (CBS), batch normalization is replaced by the feature re- IAES Int J Rob & Autom, V ol. 15, No. 3, September 2026: 544-552 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 547 nement module and the cross-spatial learning module. In the bottleneck structure, the second con v olutional layer is substituted with the Feature Renement Module, forming Bottleneck-FR. The con v olutional blocks in the spatial p yramid pooling f ast (SPPF) module are replaced with CBS-FR and CBS-CSL. Through SCEMA inte gration, spatial and channel redundancies are ef fecti v ely reduced, enabling enhanced feature representation for comple x robotic scenes. 3.1. F eatur e r enement module The feature renement module consists of a spatial reconstruction unit (SR U), spatial attention mod- ule, channel reconstruction unit (CR U), and channel attention module. As illustrated in Figure 3, the SR U performs feature separation and reconstruction by di viding the input feature map into informati v e and less- informati v e components using a scaling f actor deri v ed from group normalization (GN) [18]. Figure 3. Architecture of the SR U Gi v en an intermediate feature map X ∈ R N × C × H × W , GN is applied as, X out = GN ( X ) = γ X − µ √ σ 2 + ε + β , (1) where µ and σ denote mean and standard de viation, ε ensures numerical stability , and γ , β are learna b l e param- eters. Normalized features are mapped through a sigmoid-based g ating mechanism to generate information- a w are weights: W = Gate ( Sigmoid ( GN ( X ))) . (2) W eights abo v e a predened threshold form the informati v e mask W 1 , while remaining weights constitute the non-informati v e mask W 2 . The input feature map is decomposed as X W 1 = W 1 ⊗ X and X W 2 = W 2 ⊗ X , representing informati v e and redundant features, respecti v ely . Cross-reconstruction produces spatially rened feature map X W . As sho wn in Figure 4, the CR U further renes X W by reducing channel redundanc y through splitting, transformation, and fusion operations with compression ratio r = 2 and group size g = 2 , producing nal channel-rened features Y . Figure 4. Architecture of the CR U 3.2. Cr oss spatial lear ning module The CSL module emplo ys parallel sub-structure processing. Initially , it utilizes 1 × 1 con v olution as a comm on element. Subsequently , to inte grate spatial information across multiple scales, it positions 3 × 3 Spatial-c hannel r econstruction for ef cient multiscale attention in r obotic object (Mohammed Maiza) Evaluation Warning : The document was created with Spire.PDF for Python.
548 ❒ ISSN: 2722-2586 con v olution alongside the 1 × 1 branch. The CSL module se gments input feature map X ∈ R C × H × W into three distinct sub-feature groups, each focusing on distinct semantics. T w o paths use 1 × 1 con v olutions with one- dimensional global a v erage pooling along horizontal and v ertical directions, follo wing Coordinate Attention. The third path adopts 3 × 3 con v olution to capture multi-scale spatial features. Outputs are inte grated through matrix dot-product operations: A = Sigmoid (( W 1 × 1 ∗ X ) ⊙ ( W 3 × 3 ∗ X )) (3) where ⊙ denotes element-wise multiplication and ∗ denotes con v olution. This cross-spatial interaction enables ef fecti v e multi-scale feature fusion for robotic perception. 3.3. Integration with r obotic per ception pipelines SCEMA is designed for seamless inte gration into robotic vision stacks, includi ng R OS-based per - ception systems . The module accepts standard tensor inputs ( C × H × W ) and outputs rened feature maps with identical dimensions, enabling drop-in replacement for e xisting con v olutional blocks. W ith only 1.90M parameters and 7.4 GFLOPs, SCEMA is suitable for deplo yment on edge de vices such as NVIDIA Jetson, Raspberry Pi, and U A V onboard computers. The computational ef cienc y enables real-time operation at 30+ FPS on embedded hardw are, supporting autonomous na vig ation loops. 4. RESUL TS AND DISCUSSION This section presents comprehensi v e e v aluation of Y OLO-SCEMA across three challenging robotic scenarios: lo w-light object detection (nighttime autonomous dri ving), dense scene detection (U A V cro wd surv eillance), and comple x image structure understanding (industrial automation). Experiments were con- ducted on an NVIDIA R TX 4080 GPU with Intel Core Ultra 7 155H CPU. Hyperparameters: initial learning rate 0.01, momentum 0.937, weight decay 0.0005, input resolution 640 × 640 , training epochs 300, SGD optimizer with cosine annealing schedul er . Batch sizes: 16 for ExDark and V isDrone2019, 8 for FYP . Data augmentation: mosaic, mixup, random horizontal ip, HSV color jittering. Results represent mean ± standard de viation o v er three independent runs. Ev aluation on the ExDark dataset [19] (5,891 training, 1,472 test images, 12 cate gories, 10 illumi- nation conditions) simulates nighttime autonomous dri ving and lo w-light robotic surv eillance. T able 1 sho ws SCEMA consistently outperforms Y OLOv8n and recent attention methods including MCA [20], GAM [21], and CB AM [8]. SCEMA achie v es mAP@50 of 76.44% ± 0.28%, a 7.37% impro v ement o v er Y OLOv8n (69.07% ± 0.31%), while reducing parameters by 36.88% and computational cost by 8.64%. Compared to Y OLOv12n [15], SCEMA achie v es 0.5% higher mAP@50 using 0.66M fe wer para meters. From a robotics perspecti v e, these impro v ements translate to enhanced autonomous v ehicle safety during nighttime operation and more reliable lo w-light U A V surv eillance. T able 1. Performance comparison on ExDark dataset for lo w-light robotic perception Model mAP@50 (%) mAP@95 (%) P arams (M) GFLOPs Y OLOv8n [11] 69.07 ± 0.31 43.15 ± 0.28 3.01 8.1 MCA [20] 72.07 ± 0.29 44.36 ± 0.26 1.87 7.1 GAM [21] 73.39 ± 0.33 46.80 ± 0.30 3.94 9.5 CB AM [8] 75.58 ± 0.27 48.65 ± 0.25 3.07 8.2 Y OLOv9t [12] 74.74 ± 0.30 48.08 ± 0.27 1.97 7.6 Y OLOv10n [13] 75.47 ± 0.28 48.38 ± 0.26 2.69 8.2 Y OLOv11n [14] 75.64 ± 0.29 48.73 ± 0.27 2.58 6.3 Y OLOv12n [15] 75.94 ± 0.27 49.06 ± 0.25 2.56 6.3 SCEMA 76.44 ± 0.28 49.38 ± 0.26 1.90 7.4 Ev aluation on V isDrone2019 [22] (6,471 training, 548 v alidation, 1,580 test images) simulates U A V - based cro wd surv eillance and dense scene understanding. T able 2 sho ws SCEMA achie v es mAP@50 of 34.98% ± 0.32%, surpassing Y OLOv8n by 3.24% and outperforming PCon v [23], ECA [9], and CB AM [8] by 3.21%, 2.59%, and 2.24%, respecti v ely . Compared to Y OLOv12n, SCEMA impro v es mAP@50 by 0.33% while re- ducing parameters by 25.8%, at the cost of a modest 17.5% increase in GFLOPs. F or robotics applications, these results indicate enhanced capability for U A V -based cro wd monitoring, autonomous na vig ation in con- gested en vironments, and reliable detection of occluded objects. IAES Int J Rob & Autom, V ol. 15, No. 3, September 2026: 544-552 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 549 T able 2. Performance comparison on V isDrone2019 dataset for dense scene robotic perception Model mAP@50 (%) mAP@95 (%) P arams (M) GFLOPs Y OLOv8n [11] 31.74 ± 0.34 18.28 ± 0.30 3.01 8.1 PCon v [23] 31.77 ± 0.33 18.32 ± 0.31 1.86 7.1 ECA [9] 32.39 ± 0.31 18.53 ± 0.29 2.65 8.4 CB AM [8] 32.74 ± 0.30 18.75 ± 0.28 3.07 8.2 Y OLOv9t [12] 34.58 ± 0.29 20.15 ± 0.27 1.97 7.6 Y OLOv10n [13] 34.30 ± 0.31 20.01 ± 0.28 2.69 8.2 Y OLOv11n [14] 34.54 ± 0.30 20.24 ± 0.27 2.58 6.3 Y OLOv12n [15] 34.65 ± 0.29 20.36 ± 0.26 2.56 6.3 SCEMA 34.98 ± 0.32 20.58 ± 0.29 1.90 7.4 Ev aluation on FYP dataset [24] simulates industrial automation and manuf acturing en vironme n t s with hea vy occlusion, o v erlapping objects, and comple x structural conditions. T able 3 demonstrates SCEMA achie v es mAP@50 of 69.51% ± 0.30%, outperforming Y OLOv8n by 1.50% and surpassing EMA [25], MCA [20], and CB AM [8] by 0.45%, 0.98%, and 0.27%, respecti v ely . W ith only 1.90M parameters a n d 7.4 GFLOPs, SCEMA strik es an ef fecti v e balance for industrial robotic applications requiring structural understanding. T able 3. Performance comparison on FYP dataset for industrial automation Model mAP@50 (%) mAP@95 (%) P arams (M) GFLOPs Y OLOv8n [11] 68.01 ± 0.33 51.65 ± 0.30 3.01 8.1 EMA [25] 69.06 ± 0.31 53.16 ± 0.28 3.02 8.2 MCA [20] 68.53 ± 0.32 53.48 ± 0.29 1.87 7.1 CB AM [8] 69.24 ± 0.30 53.24 ± 0.28 3.07 8.2 Y OLOv9t [12] 68.58 ± 0.31 53.53 ± 0.29 1.97 7.6 Y OLOv10n [13] 68.67 ± 0.32 53.61 ± 0.30 2.69 8.2 Y OLOv11n [14] 68.84 ± 0.31 53.74 ± 0.28 2.58 6.3 Y OLOv12n [15] 68.95 ± 0.30 53.83 ± 0.27 2.56 6.3 SCEMA 69.51 ± 0.30 54.01 ± 0.28 1.90 7.4 Be yond standard mAP and GFLOPs, SCEMA ’ s suitability for robotics is e v aluated through addit ional metrics. Inference latenc y on NVIDIA R TX 4080: SCEMA achie v es 4.2 ms per frame (238 FPS) compared to Y OLOv8n’ s 4.8 ms (208 FPS), representing a 12.5% speed impro v ement. Estimated ener gy ef cienc y: SCEMA requires approximately 0.38 J per inference vs. 0.44 J for Y OLOv8n (13.6% reduction), critical for battery-po wered robots and U A Vs. These impro v ements enable higher update rates for robotic control loops and e xtended mission duration for autonomous systems. T able 4 analyzes each SCEMA component’ s contri- b ution across three datasets. F our v ariants are e v aluated: without SCCon v (spatial-channel reconstruction), without CSL, without CB AM, and full SCEMA. T able 4. Ablation study results across robotic perception datasets Model SCCon v CSL CB AM FYP V isDrone ExDark Y OLOv8n – – – 68.01 31.74 69.07 SCEMA w/o SCCon v – ✓ ✓ 68.69 31.95 72.95 SCEMA w/o CSL ✓ – ✓ 68.54 32.28 73.63 SCEMA w/o CB AM ✓ ✓ – 68.95 32.55 74.74 Full SCEMA ✓ ✓ ✓ 69.51 34.98 76.44 Results demonstrate full SCEMA consistently achie v es highest mAP@50. Remo ving SCCon v ca u s es performance drops of 0.82%, 3.03%, and 3.49% on FYP , V isDrone, and ExDark, respecti v ely . Remo ving CSL leads to decreases of 0.97%, 2.70%, and 2.81%. The combined impro v ement o v er baseline reaches 7.37% on ExDark, indicating strong syner gistic ef fects between spatial-channel reconstruction and cross-spatial learning. The e xperimental results demonstrate SCEMA ’ s ef fecti v eness for robotic perception. Statistical signif - icance testing using paired t-tests conrms all impro v ements are s ignicant ( p < 0 . 05 ), with p -v alues ranging from 0.003 to 0.021. Con v er gence analysis re v eals SCEMA achie v es st able training approximately 25 epochs earlier than Y OLOv8n, with nal loss v alues 0.18 lo wer , attrib uted to reduced feature redundanc y enabling more ef cient gradient o w . F or autonomous v ehicles operating at night, the 7.37% mAP impro v ement on ExDark directly addresses safety during lo w-light operation. F or drone-based surv eillance in cro wded en viron- ments, the 3.24% impro v ement on V isDrone2019 enables more reliable monitoring of occluded objects. F or Spatial-c hannel r econstruction for ef cient multiscale attention in r obotic object (Mohammed Maiza) Evaluation Warning : The document was created with Spire.PDF for Python.
550 ❒ ISSN: 2722-2586 industrial robotics requiring structural understanding, the 1.50% impro v ement on FYP f acilitates better scene interpretation and na vig ation in comple x manuf acturing en vironments. 5. CONCLUSION This paper introduced SCEMA, a lightweight and ef cient multiscale attention fusion module de- signed to enhance object detection for robotics and automation under comple x visual conditions. By emplo y- ing a parallel dual-branch architecture, SCEMA reduces spatial and channel redundanc y while strengthening multiscale feature representation through cross-spatial learning. Inte grated into Y OLOv8, Y OLO-SCEMA achie v es superior detection accurac y with signicantly fe wer parameters (1.90M vs. 3.01M) and lo wer com- putational cost (7.4 vs. 8.1 GFLOPs). Extensi v e e v aluations on ExDark, V isDrone2019, and FYP datasets demonstrate consistent outperformance of e xisting attention methods and recent Y OLO v ariants, particularly in lo w-light, dense, and structurally comple x scenes. The broader impact lies in enabling accurate real-time object detection in resource-constrained robotic en vironments, with applications in autonomous dri ving (night- time operation safety), U A V surv eillance (occluded object detection in cro wded en vironments), and industrial automation (structural underst anding in comple x manuf acturing settings). The demonstrated ability to main- tain high performance under challenging conditions while reducing computational requirements represents a signicant step to w ard practical deplo yment of dee p learning-based perception in real-w orld robotic systems. Code and pre-trained models will be made publicly a v ailable. FUNDING INFORMA TION Author state no funding is in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the C on t rib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Mohammed Maiza ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Chahira Cherif ✓ ✓ ✓ ✓ ✓ ✓ ✓ Samira Chouraqui ✓ ✓ ✓ ✓ ✓ ✓ ✓ Abdelmalik T aleb-Ahmed ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY Deri v ed data support ing the ndings of this study are a v ailable from the corresponding author upon reasonable request. REFERENCES [1] N. Ma, X. Zhang, H.-T . Zheng, and S. Sun, “Shuf eNet V2: Practical guidelines for ef cient CNN architecture de- sign, ” in Pr oceedings of the Eur opean Confer ence on Computer V ision (ECCV) , 2018, pp. 116–131, doi: 10.1007/978- 3-030-01264-9 8. IAES Int J Rob & Autom, V ol. 15, No. 3, September 2026: 544-552 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 551 [2] X. Lu, M. Sug anuma, and T . Okatani, “SBCF ormer: Lightweight netw ork capable of full-size ImageNet classica- tion at 1 FPS on single board computers, ” in Pr oceedings of the IEEE/CVF W inter Confer ence on Applications of Computer V ision (W A CV) , 2024, pp. 1123–1133, doi: 10.1109/W A CV57701.2024.00116. [3] J. Redmon and A. F arhadi, “Y OLOv3: An incremental impro v eme nt, ” arXiv:1804.02767 , 2018. [4] J. Li, Y . W en, and L. He, “SCCon v: Spatial and channel reconstruction con v olution for feature redundanc y , ” in Pr oceedings of the IEEE/CVF Confer ence on Computer V i sion and P attern Reco gnition (CVPR) , 2023, pp. 6153– 6162, doi: 10.1109/CVPR52729.2023.00596. [5] M. Sohan, T . S. Ram, and C. V . R. Reddy , “ A re vie w on Y OLOv8 and its adv ancements, ” in Pr oceedings of the Inter - national Confer ence on Data Intellig ence and Co gnitive Infor matics (ICDICI) , 2024, pp. 529–545, doi: 10.1007/978- 981-99-7962-2 39. [6] Q. Y an, Y . Feng, C. Zhang, P . W ang, and L. Zhang, “HVI: A ne w color space for lo w-light image enhancement, ” in Pr oceedings of the IEEE/CVF Confer ence on Computer V is ion and P attern Reco gnition (CVPR) , 2025, pp. 5678– 5687, doi: 10.1109/CVPR52734.2025.00533. [7] J. Hu, L. Shen, and G. Sun, “Squeeze-and-e xcitation netw orks, ” in Pr oceedings of the IEEE/CVF Confer ence on Computer V ision and P attern Reco gnition (CVPR) , 2018, pp. 7132–7141. [8] S. W oo, J. P ark, J.-Y . Lee, and I. S. Kweon, “CB AM: Con v olutional block attention module, ” in Pr oceedings of the Eur opean Confer ence on Computer V ision (ECCV) , 2018, pp. 3–19, doi: 10.1007/978-3-030-01234-2 1. [9] Q. W ang, B . W u, P . Zhu, P . Li, W . Zuo, and Q. Hu, “ECA-Net: Ef cient channel attention for deep con v olutional neural net w orks, ” in Pr oceedings of the IEEE/CVF Confer ence on Computer V ision and P attern Reco gnition (CVPR) , 2020, pp. 11531–11539, doi: 10.1109/CVPR42600.2020.01155. [10] J. T erv en, D.-M. C ´ ordo v a-Esparza, and J.-A. Romero-Gonz ´ alez, “ A comprehensi v e re vie w of Y OLO architectures in computer vision: From Y OLOv1 to Y OLOv8 and Y OLO-N AS, ” Mac hine Learning and Knowledg e Extr action , v ol. 5, no. 4, pp. 1680–1716, 2023, doi: 10.3390/mak e5040083. [11] “Y OLOv8: A state-of-the-art object detect ion model, ” GitHub , 2023. https://roboo w .com/model/yolo v8. [12] C.-Y . W ang, I.-H. Y eh, and H.-Y . M. Liao, “Y OLOv9: Learning what you w ant to learn using programmable gra- dient i nformation, ” in Pr oceedings of the Eur opean Confer ence on Computer V ision (ECCV) , ser . Lecture Notes in Computer Science, v ol. 15089, 2025, pp. 1–18, doi: 10.1007/978-3-031-72751-1 1. [13] A. W ang et al. , “Y OLOv10: Real-time end-to-end object detection, ” in Pr oceedings of the International Confer ence on Neur al Information Pr ocessing Systems (NeurIPS) , v ol. 38, 2024. [14] C. W ang, X. Song, J. W ang, and Y . Zhang, “ An impro v ed Y OLOv11 algorithm for small object detection in U A V images, ” Signal, Ima g e and V ideo Pr ocessing , v ol. 19, p. 507, 2025, doi: 10.1007/s11760-025-04157-w . [15] Y . Qiao, Y . W ei, and K. W ang, “F ast-Y OLOv12: An attention-guided lightweight netw ork for real-time steel surf ace defect detection, ” J ournal of Real-T ime Ima g e Pr ocessing , v ol. 23, p. 9, 2026, doi: 10.1007/s11554-025-01807-7. [16] T .-Y . Lin, P . Doll ´ ar , R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature p yramid netw orks for object detec- tion, ” in Pr oceedings of the IEEE/CVF Confer ence on Com puter V ision and P attern Reco gnition (CVPR) , 2017, pp. 2117–2125, doi: 10.1109/CVPR.2017.106. [17] M. Maiza, C. Cherif, S. Chouraqui, and A. T aleb-Ahmed, “Enhanced U A V na vig ation in cluttered en vironments via a BS-CY OLOv5 multi-sensor approach, ” Disco ver V ehicles , v ol. 2, p. 3, 2026. [18] Y . W u and K. He, “Group normalization, ” in Pr oceedings of the Eur opean Confer ence on Computer V ision (ECCV) , 2018, pp. 3–19, doi: 10.1007/978-3-030-01261-8 1. [19] Y . P . Loh and C. S. Chan, “Getting to kno w lo w-light images with the e xclusi v ely dark dataset, ” Computer V ision and Ima g e Under standing , v ol. 178, pp. 30–42, 2019, doi: 10.1016/j.cviu.2018.10.010. [20] Y . Y u, Y . Zhang, Z. Cheng, Z. Song, and J. T ang, “MCA: Multidimensional col laborati v e attention in deep con- v olutional neural netw orks for image recognition, ” Engineering Applications of Articial Intellig ence , v ol. 126, p. 107079, 2023, doi: 10.1016/j.eng appai.2023.107079. [21] Z.-Y . Ni and J.-C. W ang, “ A global attention mechanism-based Ef cientNet model for road pa v em ent-type identi- cation, ” International J ournal of Computational Intel lig ence Systems , v ol. 18, p. 97, 2025, doi: 10.1007/s44196-025- 00842-3. [22] D. Du et al. , “V isDrone-DET2019: The vision meets drone object detection in image challenge results, ” in Pr oceed- ings of the IEEE /CVF International Confer ence on Computer V ision W orkshops (ICCVW) , 2019, pp. 213–226, doi: 10.1007/978-3-030-11021-5 27. [23] S. P ark, Y .-J. Y eo, and Y .-G. Shin, “PCon v: Simple yet ef fecti v e con v olutional layer for generati v e adv ersarial net- w ork, ” Neur al Computing and Applications , v ol. 34, pp. 7113–7124, 2022, doi: 10.1007/s00521-021-06846-2. [24] “FYPproject dataset, ” Roboow Univer se , Oct. 2022. https://uni v erse.roboo w .com/muaz-ahmed/fypproject. [25] D. Ouyang et al. , “Ef c ient multi-scale attention module with cross-spatial learning, ” in Pr oceedings of the IEEE International Confer ence on Acoustics, Speec h and Signal Pr ocessing ( ICASSP) , 2023, pp. 1–5, doi: 10.1109/ICASSP49357.2023.10096516. Spatial-c hannel r econstruction for ef cient multiscale attention in r obotic object (Mohammed Maiza) Evaluation Warning : The document was created with Spire.PDF for Python.
552 ❒ ISSN: 2722-2586 BIOGRAPHIES OF A UTHORS Mohammed Maiza is a research scientist in the Computer Science Department at the Uni v ersity of Oran 1 Ahmed Ben Bella, Algeria. He holds an M.Sc. and a Ph.D. in computer science from the Uni v ersity of Sciences and T echnology of Oran. His research focuses on bioinformatics, computational biology , and machine learning, with emphasis on high-dimensional gene selection, mi- croarray data anal ysis, and h ybrid e v olutionary algorithms for cancer classication. He collaborates with medical r esearchers to de v elop computational methods with clinical applications in genomics and precision medicine. He can be contacted at email: maiza.mohammed@uni v-oran1.dz. Chahira Cherif is a research scientist in computer science at the Uni v ersity of Oran 1 Ahmed Ben Bella, Algeria , where she earned her engineering, master ’ s , and Ph.D. de grees. Her research focus es on arti cial intelligence, b usiness process ma nagement, decision support systems, and b usiness rules modeling, emphasizing the inte gration of AI into b usiness process optimization and decision-making. She contrib utes to national and international research projects, supervises graduate students, and her current w ork centers on intelligent decision support systems using ma- chine learning, particularly in healthcare and industrial conte xts. She can be contacted at email: cherif.chahira@uni v-oran1.dz. Samira Chouraqui is a full professor in computer science at the Uni v ersity of Sciences and T echnology of Oran (UST O MB), Algeria. She earned an engineering diploma in electrical engineering from UST O Oran, an MSc in satellite communication from the Uni v ersity of Surre y , UK, and a Ph.D. in computer science from UST O Oran. Her research spans computer vision, AI, U A Vs, pattern recognition, a nd machine learning for remote sensing, with current focus on intelligent vision systems, U A V autonomous na vig ation, and AI-dri v en solutions for en vironmental monitoring and precision agriculture. She has supervised numerous graduate students and led se v eral national research projects. She can be contacted at email: samira.chouraqui@uni v-usto.dz. Abdelmalik T aleb-Ahmed is a full professor at the Polytechnic Uni v ersity of Hauts-de- France, France, af liated with IEMN. His research co v ers computer vision, machine learning, pattern recognition, image s e gmentation, classication, and data fusion, with applications in biometrics, video surv eillance, autonomous dri ving, and medical imaging. He has co-authored o v er 85 peer - re vie wed papers, supervised 30 graduate students, and led national and Europe an research projects. His current w ork focuses on deep learning for medical image analysis, multimodal data fusion for autonomous systems, and adv anced pattern recognition for security applications. He can be contacted at email: abdelmalik.taleb-ahmed@uphf.fr . IAES Int J Rob & Autom, V ol. 15, No. 3, September 2026: 544-552 Evaluation Warning : The document was created with Spire.PDF for Python.