IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 4, August 2026, pp. 3326 ∼ 3341 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i4.pp3326-3341 ❒ 3326 A transfer lear ning-based appr oach f or automatic monument detection Santosh Giri 1 , J ebish Purbey 1 , Sunil Adhikari 1 , P aarit P okhar el 1 , Bab u R. Dawadi 1 , Bipun Man P ati 2 , Atsushi Ito 3 , Sushant Chalise 1 , Sanji v an Satyal 1 1 Department of Electronics and Computer Engineering, Pulcho wk Campus, Institute of Engineering, T ribhuv an Uni v ersity , Lalitpur , Nepal 2 Adv anced Colle ge of Engineering and Management, T ribhuv an Uni v ersity , Kathmandu, Nepal 3 Chuo Uni v ersity , Hachioji, Japan Article Inf o Article history: Recei v ed No v 20, 2024 Re vised Jun 19, 2026 Accepted Jul 9, 2026 K eyw ords: Architectural heritage Articial intelligence Con v olutional neural netw ork Deep learning MobileNet Monument recognition T ransfer learning ABSTRA CT Architectural heritage connects us to the cultural achie v ements of past ci vilizations. P atan Durbar Square in Nepal is home to man y such structures, yet identifying them remains a challenge for tourists. This paper presents an automated monument recognition system b uilt on the backbone of con v olutional neural netw orks (CNNs). A dataset of 1832 images of 9 important monuments from P atan w as created, and b uild a detection system with MobileNetV2, a light-weight CNN, to detect monuments in P atan Durbar Square with a near -perfect F1 score of 98.94%. The approach utilizes transfer learning to adapt the model to local architectural styles. A detailed ablation study is performed to determine the optimal netw ork design and augmentation strate gies. Class-wise performance is further analyzed to v erify rob ustness ag ainst visual occlusion and similarity . Finally , the model is deplo yed as a mobile application using Flutter and the F astAPI frame w ork. This w ork demonstrates the viability of lightweight CNNs for real-time cultural heritage preserv ation. This is an open access article under the CC BY -SA license . Corresponding A uthor: Santosh Giri Department of Electronics and Computer Engineering, Pulcho wk Campus, Institute of Engineering T ribhuv an Uni v ersity Lalitpur , Nepal Email: santoshgiri@pcampus.edu.np 1. INTR ODUCTION Cultural heritage stands as a tangible testament to human history and artistic e v olution. The Kathmandu V alle y serv es as a prime e xample, hosting se v en UNESCO W orld Heritage sites within a small radius [1]. This high density of historical sites mak es it a major hub for cultural tourism in South Asia. Recent data from the Nepal T ourism Board highlights a strong post-pandemic reco v ery in visitor numbers [2], [3]. Ho we v er , the sheer density of monuments often o v erwhelms visitors who lack deep historical kno wledge. Most tourists f ace signicant challenges in identifying specic shrines and temples among the clustered architecture. Professional local guides can be e xpensi v e or una v ailable during peak seasons. Furthermore, static guidebooks often f ail to pro vide instant visual identication from a specic vie wpoint. This g ap creates a need for digital tools that can of fer real-time information without human interv ention. A mobile-based solution can democratize access t o historical kno wledge for all visitors. T o address this challenge, an end-to-end system is proposed that utilizes deep learning to detect and recognize J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3327 monuments in real-time. The system is de v eloped by creating a di v erse dataset of 1,832 images that captures nine distinct monuments under v arying lighting conditions and angles. A lightweight con v olutional neural netw ork (CNN) model is trained on this data, utilizing transfer learning to achie v e high accurac y while maintaining the lo w latenc y required for mobile de vices. This system allo ws users to simply point their camera at a structure to recei v e immediate histori cal conte xt. An earlier v ersion of this w ork w as presented at the International Symposium on Computing and Netw orking W orkshops (CAND AR W) [4]. This m anuscript e xtends that w ork by pro viding deeper analysis, e xpanded methodological justication, and additional discussion on rob ustness and generalization. 2. LITERA TURE REVIEW Image classication is a foundational problem of computer vi sion, which has as its main objecti v e to label an image based on its visual s. The emer gence of deep learning and, to a greater e xtent, CNNs [5], [6] has greatly helped in vision tasks, especially for applications in object detection, medical imaging, and self-dri ving. CNNs are a brack et of deep learning architectures that automatically and adapti v ely learn spatial features of input from lo w-le v el edges to high-le v el patterns. CNNs can identify patterns directly from ra w pix el data, signicantly impro ving the ef cienc y and accurac y of image recognition tasks [7]. The foundational w orks in [8]–[10] made substantial contrib utions to progress in computer vision and image classication, particularly for emplo ying v ery deep neural netw orks. Building upon this, Saini et al. [11] pioneered the use of CNN techniques for autonomous recognition of Indian cultural heritage sites, opening fresh possibilities for research in monument detection and recognition tasks. T ransfer learning enables adaptation of pre-trained models to ne w tasks with limited data. By le v eraging features learned on lar ge datasets lik e ImageNet [12], transfer learning reduces computational requirements and training time while impro ving generalization [13], [14]. This approach pro v es particularly v aluable for monument recognition, where training data may be limited. Recent adv ances in monument recognition ha v e le v eraged v arious CNN architectures and deep learning techni qu e s. P aul et al. [15] conducted an e xtensi v e analysis of machine lear ning and deep learning approaches for Indian monument classication, demonstrating the ef fecti v eness of transfer learning and CNN-based methods. Hesham et al. [16] compar ed deep learning v ersus traditional machine learning approaches for monument recognition, establishing the superiority of CNN-based techniques. Shamo v and Shelest [17] appli ed CNNs specically for the recognition of to wer lighthouses, demonstrating the adaptability of deep learning techniques across v arious domains. Jaisw al et al. [18] studied identifying monuments through aerial images using you only look once v ersion 5 (Y OLOv5) [19] and visual geometry group 16 (V GG-16) [8] as the base model. Kalliatakis and T riantafyllidis [20] e xplored utilization of graph-based visual salienc y for monument recognition from images. T ermritthikun and Muneesa w ang [21] de v eloped Naresuan Uni v ersity-lightweight netw ork (NU-LiteNet), a no v el lightweight CNN inspired by SqueezeNet [22], achie ving competiti v e accurac y measures on monument datasets. Using top-1 and top-5 accurac y measures, NU-LiteNet ac hie v ed greater accurac y , e v en surpassing GoogLeNet [23] for the Sing apore and P aris datasets [24]. Razali et al. [25] de v eloped a landmark recognition model for smart tourism appl ications using lightweight deep learning combined with linear discriminant analysis, achie ving 100% classication accurac y on the Uni v ersiti Malaysia Sabah-scene (UMS-Scene) dataset and 94.26% on the Scene-15 dataset using the Ef cientNet-CNN architecture. Additionally , t he inte gration of data augmentation techniques, including random ip, rotation, translation, zoom, contrast, hue, brightness, and saturation adjustments, has pro v en ef fecti v e in impro ving rob ustness [26]. W ang et al. [27] applied an impro v ed Inception-V3 model for ancient architecture recognition, incorporating dropout layers and transfer learning to address o v ertting in small monument datasets. Mobile deplo yment of monument recognition systems has g ained prominence, with researchers focusing on de v eloping lightweight CNN architectures suitable for real-time heritage site identication on mobile de vices, enhancing accessibility for tourism and cultural preserv ation initiati v es [28], [29]. 3. MA TERIALS AND METHODS This section outlines the theoretical frame w ork, data collection, and methodology for the m onument recognition application. The methodology for monument recognit ion is subdi vided into se v eral steps. Each step is described in the follo wing subsections. A tr ansfer learning-based appr oac h for automatic monument detection (Santosh Giri) Evaluation Warning : The document was created with Spire.PDF for Python.
3328 ❒ ISSN: 2252-8938 3.1. Dataset cr eation T o train the model, the dataset w as compiled by clicking pictures of cultural heritages around the P atan Durbar Square area. The dataset includes a total of 1,832 images belonging t o 9 dif ferent classes as sho wn in T able 1. Figure 1 sho ws a bat ch of the dataset. The images were click ed with 4 dif ferent mobile de vices with v arying dimensi ons to ensure that a particular de vice doesn’ t af fect the result. Ho we v er , for training, the images were resized to 224×224 during the pre-processing steps. In the preprocessing phase, both data augmentation and normalization were applied. Initially , each image w as rotated 270 ◦ counter -clockwise during the dataset loading phase to correct their orientation. This adjustment w as necessary because the images, which were captured in portrait mode (i.e., width < height), were interpreted as landscape mode by def ault during loading, causing an unw anted 90 ◦ counter -clockwise rotation. T o address this, the images were rotated back to their correct orientation. The dataset w as split into 70 : 15 : 15 ratios for training, v alidation, and testing respecti v ely . F or data augmentation, random rotation ( 0 . 2 , 0 . 4 and 0 . 7 radians), random brightening ( 40 %), random contrasting ( 40 % and 60 %), and random horizontal and v ertical ipping were applied. This process created ne w images by rotating, adj u s ting intensity , and ipping the originals, which increased the dataset size and introduced v ariability . This v ariability helps the model adapt to dif ferent angles and perspecti v es of the monuments. Normalization w as achie v ed using MobileNetV2 [30] preprocessing, which scales the pix el v alues within the range [ − 1 , 1] . The data augmentation is only acti v ated during the training phase by def ault. T able 1. Image distrib ution for each class for each phase Monument T rain set V alidation set T est set T otal Bhimsen T emple 191 40 42 273 Char Narayan T emple 156 33 35 224 Dhunge y Dhara 114 24 26 164 Garuda Pillar 70 15 15 100 Harishankara T emple 140 30 30 200 Krishna Mandir 172 36 38 246 Narayan T emple 169 36 37 242 Octagonal Ch yasing De v al 130 27 29 186 V ishw anath T emple 137 29 31 197 T otal 1,279 270 283 1,832 Figure 1. A sample batch of the dataset after e v ery image has been normalized and resized to 224×224 3.2. Monument class selection and dataset r epr esentati v eness Nine monument classes were selected from P atan Durbar Square. Bhimsen T emple, Char Narayan T emple, Dhunge y Dhara, Garuda Pillar , Harishankara T emple, Krishna Mandir , Narayan T e mple, Octagonal Ch yasing De v al, and V ishw anath T emple. The selection w as guided by the n e ed to balance dataset representati v eness with the practical constraints of data collection and annotation. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3326–3341 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3329 3.2.1. Selection rationale T o identify appropriate monuments for the task, t he help of a local tourist guide w as tak en. The selected heritage sites ha v e signicant cultural and historical v alue. The goal w as to include monuments that are cul turally important b ut also widely recognized by visitors. P atan Durbar Square is one of the most important historical sites in Nepal that reects the artistic nuances of the Malla era. It is a designated UNESCO W orld Heritage site that sits at the heart of Lalitpur city . T emples, shrines, and palace s embodying Ne w ari craftsmanship and traditions impart cultural life to e v ery corner of the square. The selection of monuments captures this theme and the beauty of P atan Durbar Square. Input features were also considered in the selection of monuments. The selection consists of cultural heritages that are a blend of distincti v e and similar architectures. Some structures are multi-tiered pagodas typical of traditional Ne w ari T emples, while still ha ving vi vid dif ferences among them. Krishna Mandir , which is a stone construction and has a v ery dif ferent f acade, is completely dif ferent from the Ne w ari-traditional temples. Each of these monuments also has pillars, w ater spouts, and shrines, which int roduces morphological v ariety . Such a dataset w as created to impro v e the generalization of the model. Finally , monuments that are wi dely popular among tourists were considered. The selection of monuments is among the most popular hotspots within P atan Durbar Square. People frequently visit them and photograph them, especially the Krishna Mandir . As a result, the model de v eloped has a v ery high de gree of practicality . All the monuments were captured under v arying light conditions, across days, to ensure rob ust model training. 3.2.2. Repr esentati v eness The selection of monuments co v ers a wide range of religious, cultural, and historical conte xts pres ent in P atan Durbar Square. P atan Durbar is home to more than 100 monuments, and thus, our dataset is not an e xhausti v e collection of monuments within this area. Ho we v er , the selection reects the k e y architectural and cultural characteristics of the site. Furthermore, the y are 9 of the most visited and referred sites within the area. As such, the y can be considered a good representation of the artistic landscape of P atan Durbar Square. Y et, the dataset is only representati v e of a single heritage zone, while Kathmandu V alle y itself contains tw o more Durbar areas with rich cultural conte xts. Future w ork should e xte nd be yond P atan to other sites wi thin the V alle y , and hopefully o v er the country , and impro v e the generalization capability of m odels. Ho we v er , for the current scope of the w ork, these nine classes pro vide an adequate base for analyzing the capability of automated monument recognition in the conte xt of edge de vices. 3.3. Model de v elopment The architecture of the model is illustrated in Figure 2. In this gure, MobilenetV2 is used as the backbone for the model. T ransfer learning is emplo yed t o design the classication netw ork based on the MobileNetV2 backbone, which outputs a feature map of size 7 × 7 × 1 , 280 . MobileNetV2 is a lightweight and ef cient CNN designed for lo w-resource vision applications. It introduces se v eral k e y enhancements, most notably the use of in v erted residuals and linear bottlenecks. It utilizes depth-wise separabl e con v olutions that notably reduce the number of calculations required for performing con v olutions, resulting in a lightweight, lo w-parameter -count netw ork. In v erted residuals enable the netw ork to preserv e feature maps through nonlinearities, which impro v es the representational capacity of the model. Figure 2. Architectural diagram of the model for monument recognition constructed using V isualk eras [31] A tr ansfer learning-based appr oac h for automatic monument detection (Santosh Giri) Evaluation Warning : The document was created with Spire.PDF for Python.
3330 ❒ ISSN: 2252-8938 Depthwise separable con v olution is an approach in computer vision that reduces both comput ational cost and model size without compromising performance. It splits the con v olution operation into tw o distinct steps: depthwise con v olution accompanie d by pointwise con v olution. A separate lter is applied to each input channel indi vidually during depthwise con v ol ution. Normally , with n input channels, a standard con v olution uses a lter with depth n , b ut in depthwise con v olution, n lters each with a depth of 1 are used, producing an output with a depth of 1 per lter . This operation is succeeded by a pointwise con v olution, where 1 × 1 lters of depth n are applied to the depthwise con v olution’ s output. The number of these 1 × 1 lters determines the nal output depth. Depthwise separable con v olution signicantly reduces the number of computations, leading to f aster inference times. F or an input feature map X with height H , width W , and depth D ( H × W × D ), depthwise con v olution applies a separate k ernel indi vidually to each channel. Let K d represent the depthwise lter with dimensions k h × k w × 1 , where k h is the height and k w is the width of the lter . F or each input channel d (where d ranges from 1 to D ), the depthwise con v olution output Y d can be dened as in (1). Y d ( i, j ) = k w − 1 X n =0 k h − 1 X m =0 K d ( m, n ) · X d ( i + m, j + n ) (1) Where, i and j are spatial indices and, K d is the depthwise lter applied to the d -the channel of the input. After performing depthwise con v olution, the pointwise con v olution k ernel is applied to the output across dif ferent channels. The pointwise con v olution uses a 1 × 1 lter K p with dimensions 1 × 1 × D × M , where M is the desired number of output channels. The output Z m for the m -th output channel can be e xpressed as in (2). Z m ( i, j ) = D X d =1 K p ( d, m ) · Y d ( i, j ) (2) Where K p ( d, m ) is the pointwise lter coef cient connecting the d -th input channel to the m -th output channel, and Y d ( i, j ) is the output of depthwise con v olution for channel d . The MobileNetV2 backbone is succeeded by a global a v erage pooling (GAP) layer and an imm ediate dropout layer . GAP reduces 7 × 7 × 1 , 280 feature maps to a 1 , 280 -dimensional v ector by a v eraging the v alues in each 7 × 7 feature map. This reduces the dimension and computational comple xity . A dropout layer of probability p , is then emplo yed to mitig ate o v ertting by setting a fraction of the neurons’ output to zero during training wi th probability p . This combination ensures that the model generalizes well and maintains high performance on no v el data. Ne xt, a fully connected block is introduced to w ork as the classier for the model. The a rchitecture of the fully connected block is v aried to determine the optimal netw ork design. The fully connected netw ork (FCN) architectures under consideration are: i) simple FCN: a single-layer FCN with 9 neurons; ii) tw o-layer FCN: a double-layer FCN, ha ving 18 neurons follo wed by 9 neurons; and i ii) three-layer FCN: a FCN with three layers, consisting of 36, 18, and 9 neurons in the rst, second, and nal layers, respecti v ely . Additionally , each architecture is tested with and without dropout to e v aluate its impact on performance. The detailed results of these e xperiments, including the impact of dropout, are discussed in the results section. 3.4. Model choice justication MobileNetV2 [25] is a lightweight CNN model that is specically b uilt for edge-case usages. In this conte xt, MobileNetV2 w as chosen because of its high accurac y and ef cienc y in public benchmarks. This is important for the system as it aims for real-w orld deplo yment in a resource-constrained en vironment. 3.4.1. Rationale f or MobileNetV2 MobileNetV2 uses depthwise separable con v olutions for con v olution operations. This reduces parameters and oating-point operations while still maintaining high performance. As a result, the model is comparati v el y v ery small in weight and is a solid option for mobile deplo yment. Since practical u s ability is a long-term goal of this research, model ef cienc y w as weighted alongside accurac y in the design decisions. MobileNetV2 also performs v ery well in transfer learning. Since the dataset of 1,832 images is v ery modest compared to standard model training datasets, transfer learning w as a necessity . T ransfer learning has also pro v en to be a much better option than training from scratch, e v en in cases of huge dataset a v ailability . Int J Artif Intell, V ol. 15, No. 4, August 2026: 3326–3341 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3331 By using MobileNetV2, the weights of the model that were already trained on 1,000 classes of ImageNet were le v eraged [24]. This allo wed for f aster con v er gence and high accurac y right of f the bat during training. As a result, the model achie v ed 99% accurac y in the e xperiments. 3.4.2. Consideration of alter nati v e ar chitectur es Other architectures that are widely used for detection tasks were also considered. Ef cientNet [32] models of fer v ery strong performance, abo v e that of MobileNet. Ho we v er , the y are hea vier and thus require higher comput ation resources than MobileNet. This decreases their practicality and mak es them less viable for edge-case uses. Y OLO-based models [19] are one of the most widely used models in computer vision. The y of fer the highest performance in comple x scenarios and sub-second detection, making them useful for detection within videos too. Ho we v er , since monument classication is a much simpler task where a monument dominates the image, adding a full detection pipeline complicates the system and introduces the risk of o v ertting. As MobileNetV2 w as pro viding similar performance with a simpler pipeline, it w as selected for the model. 3.4.3. Design trade-off perspecti v e The model selection reects a deliberate trade-of f that prioritizes practicality and generalization o v er higher numbers. A lightweight model reduces the risk of o v ertting and is easier to deplo y in lo w-resource de vices such as smartphones. These are practical adv antages that speed up the model de v elopment as well as the deplo yment time. Still, other lightweight models with similar usability are recognized and discussed in subsection 3.4.2. Ho we v er , within the scope of this w ork, MobileNetV2 stands out as a solid option for the backbone of the recognition pipeline. 3.5. T raining and e v aluation The functioning of all models w as assessed on the monument recognition dataset using accurac y and loss as performance metrics. The total training epochs for each model w as 80 epochs, 50 epochs of classication-head training with a learning rate of 0 . 001 , follo wed by 30 additional epochs of ne-tuning with the learning rate decreased to 0 . 0001 . During ne-tuning, the last 29 layers of the base model were unfrozen. This training process w as repeated for 3 turns for each model to account for v ariability introduced by stochastic f actors such as random initialization, data shuf ing, and batch processing. The accurac y metrics for the last 5 epochs are also calculated to assess the model’ s stability during the nal phase of ne-tuning. The most rob ust model is identied with a focus on minimizing o v ertting. After selecting the best model, a h yperparameter search is conducted to maximize performance after t raining. The h yperparameters under consideration are listed in the T able 2. The base model i.e., MobilenetV2 consists of a total of 154 layers including the input layer . The main con v olutional block with residual connection consists of 9 layers. The h yperparameter space for the number of layers to unfreeze is chosen in such a w ay that each step-increase unfreezes a con v olutional block in the base model. Random search is used to nd the best h yperparameter combinations for both initial training and ne-tuning. Early stopping with a patience v alue of 5 is also emplo yed during the s earch, with maximum trials set to 75 for initial training and 40 for ne-tuning. The best h yperparameter combination is then used to train the model and e v aluate its performance on the test set. The performance metrics used are recall, precision, and F1-score as dened in [33]. T able 2. Hyperparameters search space Phase Hyperparameter V alue Initial training Learning rate 0.01 to 0.0001 (Sampling = log) Dropout rate 0.2, 0.3, 0.4, 0.5 and 0.6 No. of epochs 20, 30, and 40 Batch size 0.2, 16, 32, and 64 Re gularization f actor (L2) 0.1 to 0.001 (Sampling = log) Optimizer Adam, root mean square propag ation (RMSprop) and Stochastic gradient descent (SGD) (with momentum) Fine tuning Learning rate 0.001 to 0.000001 (Sampling = log) No. of layers to unfreeze 29, 38, 47, 56, and 65 No. of epochs 30, 40, and 50 A tr ansfer learning-based appr oac h for automatic monument detection (Santosh Giri) Evaluation Warning : The document was created with Spire.PDF for Python.
3332 ❒ ISSN: 2252-8938 3.6. Data augmentation rationale Data augmentation h e lps impro v e the rob ustness of deep learning models when the dataset is small. Since the dataset is relati v ely small, data augmentation w as used in the training pipeline. This ef fecti v ely increases the di v ersity within the same sample and reduces the risk of o v ertting. 3.6.1. Selected augmentation strategies The augmentation pipeline consisted of se v eral dif ferent image transformation operations. These include rotation, brightness, contrast adjustments, and horizontal ipping. These transformations were chosen as the y reect the most commonly occurring scenarios where people click images. Rotation represents the slight vie wpoint changes that are caused when handheld cameras get tilted while clicking images. This is also one of the most common forms of aberrations in images. C asual photographers rarely maintain a perfectly aligned camera, and these slight angular v ariations are common in click ed images. Ha ving datasets that account for these v ariations helps the model become rob ust when f acing such edge cases in practical usage. Brightness and contrast adjustments are incorporated to account for images click ed under dif ferent lighting conditions. Depending upon the time of day and weather , images captured of monuments can dif fer signicantly . While the original dataset itself contains images click ed during dif ferent times of the day , augmentation using brightness and contrast modications complements the training further . This allo ws the model to learn features that are less dependent on specic lighting conditions. Horizontal ipping is used to increase the visual di v ersity of the samples. Man y monuments e xhibit approximate bilateral symmetry and maintain identity e v en when mirrored. This also simulates cases where images are not ipped after capturing with a front-f acing camera. 3.6.2. P arameter selection considerations The augmentation parameters are k ept within a moderate range that simulates the real-life a berration scenarios. This also pre v ents generating unrealistic samples, which only consumes training time b ut doesn’ t pro vide an y g ains. The intention with augmentation w as to complement our already di v erse dataset with plausible photographic v ariations. Extreme transformation can lead to de graded model learning, as it can distort the monument completely and shift the model’ s weight to w ards recognizing these distorted samples. 3.6.3. Consider ed b ut excluded techniques Other widely used augmentation techniques, such as random cropping and noise injection. Ho we v er , the y were not included in the augmentation pipeline, as the y are not indicati v e of practical photographic v ariations. Random cropping has the risk of remo ving the k e y architectural feature within the image, and may lead to label ambiguity . Similarly , injecting articial noise could lead to the model trying to learn the features of the noise instead of the monuments. As the initial dataset already contains realistic v ariations of noise and cropping, the models tr ained on the dataset are well-equipped to handle such aberrations in practical use. 3.6.4. Role in generalization The augmentation strate gy w as designed to introduce realism and balance di v ersity . By e xposing the model to the controlled v ariations of the same samples, it is encouraged to learn in v ariant features. This results in impro v ed generalization and rob ust performance across conditions. 3.7. De v elopment After e v aluating the best model, it is deplo yed as a smartphone application for real-time monument recognition. F ast application programming interf ace (F astAPI) is used to de v elop the back end serv er , which handles model inference requests. F astAPI allo ws asynchronous operations for seamless communication between the smartphone application and the back end serv er . F or the mobile application, the Flutter frame w ork is used, which allo ws us to b uild a cross-platform app from a single codebase. T o manage h yperte xt transfer protocol (HTTP) requests and interactions with the F astAPI back end, Dio is emplo yed, a po werful HTTP client for Flutter . 4. RESUL TS AND DISCUSSION This section presents a comprehensi v e analysis of the model training and e v aluation outcomes, including ablation studies to v alidate architectural choices and the ef fecti v eness of re gularization techniques. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3326–3341 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3333 Dif ferent h yperparameter congurations are systematically e v aluated to identify the optimal settings for nal model training, ensuring rob ust performance across multiple v alidation metrics. Additionally , the model’ s genera lization capabilities are assessed through rigorous testing on an independent test set, e xamining both classication accurac y and error patterns to v alidate the practical utility of the monument recognition frame w ork. 4.1. Ablation study Fi v e distinct models were e xperimented with to e xamine ho w dif ferent netw ork designs and dropout rates impact performance. T able 3 summarizes the ndings, sho wing a v erage accurac y and a v erage loss metrics for both training and v alidation phases for 3 total training runs. The single fully connected architecture (9 neurons) achie v ed the highest training and v alidation accurac y at 99 . 77 % and 96 . 42 % respecti v ely . When comple xity w as added with tw o fully connected layers ( 18 and 9 neurons) without dropout, a signicant drop in v alidation accurac y and an increase in v alidation loss were observ ed. This indicates that more comple x models without dropout might o v ert the training data. The three-layer model attained an accurac y of 99 . 40 % for training b ut performed poorly for the v alidation dataset ( 92 . 84 %) compared to simpler architectures, sho wing signs of o v ertting. Incorporating dropout ( 0 . 2 ) into the tw o-layer and three-layer models reduced training accurac y b ut k ept v alidation accurac y consistent. This sho ws that dropout helps in reducing o v ertting by impro ving generalization, e v en though it af fects training accurac y . The most comple x conguration with three fully connected layers and dropout ( 0 . 2 ) pro vided the second highest v alidation accurac y at 94 . 44 % and impro v ed performance during ne-tuning. It should be noted that the training accurac y sho wn for dropout models will be lesser than their actual training accurac y as dropout will reduce the accurac y during training and not during e v aluation. The a v erage accurac y metri cs of the last 5 epochs sho w that the single FC architecture and 36 − 18 − 9 dropout had the most stable training during the last phases. Gi v en the comparable performance between the tw o models, Occam’ s Razor [34] principle is used to select the simpler model. Occam’ s Razor principle suggests that among competing h ypotheses that e xplain the data equally well, the simplest one should be selected. The single fully connected layer model thus emer ged as the most rob ust conguration, balancing high accurac y and minimal o v ertting with the simplest architecture design. T able 3. A v erage training results for all v e fully connected architectures for 3 turns of training Model Initial-training (50 epochs) Fine-tuning (30 epochs) A v erage of last 5 epochs T raining V alidation T raining V alidation T raining V alidation Accurac y Los s Accurac y Los s Accurac y Loss Accurac y Loss Accurac y Accurac y 9 0.9523 0.1461 0.8543 0.4272 0.9977 0.0069 0.9642 0.0874 0.9971 0.9514 18-9 0.9510 0.1433 0.8556 0.4185 0.9974 0.0122 0.9111 0.2858 0.9968 0.9188 18-9 dropout 0.8027 0.5303 0.8222 0.5423 0.9322 0.1827 0.9222 0.2186 0.9298 0.9081 36-18 0.9453 0.1541 0.8469 0.4247 0.9940 0.0159 0.9284 0.2184 0.9944 0.9163 36-18-9 dropout 0.8548 0.4267 0.8469 0.4987 0.9487 0.1485 0.9444 0.2111 0.9484 0.9321 4.2. Hyper parameter sear ch and training The h yperparameters listed in T able 2 were searched for using random search. The best v alues for each h yperparameter are listed in T able 4. The training of the single-layer ( 9 neurons) model w as conducted for a total of 80 epochs. The h yperparameters used were the ones that g a v e the best v alidation accurac y during the h yperparameter search. F or the ne-tuning, the l ast 47 layers of the base model were unfrozen. The training curv es of the complete process are sho wn in Figure 3. The pink dotted line in this gure indicates the start of ne-tuning with the last 47 layers of MobileNetV2 unfrozen. The nal training accurac y of the model is 100 . 00% with a loss of 0 . 0345 . The model demonstrated e xceptional performance on the v alidation set as well, with a loss of 0 . 0470 and a near -perfect accurac y of 99 . 63 %. The model’ s v alidation accurac y at the end of the 30th epoch, i.e., before ne-tuning w as 82 . 22 %. These ndings sho w that the model’ s performance and generalization impro v ed considerably during the ne-tuning phase. The ne gligible dif ference between the accurac y metrics also indicates that the model is not o v ertting. A tr ansfer learning-based appr oac h for automatic monument detection (Santosh Giri) Evaluation Warning : The document was created with Spire.PDF for Python.
3334 ❒ ISSN: 2252-8938 T able 4. Hyperparameters best v alues Phase Hyperparameter Best v alue Initial training Learning rate 0.001 Dropout rate 0.2 No. of epochs 30 Batch size 32 Re gularization f actor (L2) 0.006285 Optimizer Adam Fine tuning Learning rate 0.000193 No. of layers to unfreeze 47 No. of epochs 50 Figure 3. T raining curv es for the best model 4.3. T est set e v aluation A nal e v aluation of the model w as conducted on the test set consisting of 283 images across the 9 classes with distrib ution as in T able 1. The model achie v ed a splendid accurac y of 98 . 94 % with minimal loss of 0 . 0810 . The performance across the classes w as splendid as perfect F1-score, recall, and precision scores were achie v ed in 4 classes. The performance metrics of indi vidual classes along with the a v erage are listed in T able 5. Out of 283 images, only 3 instances of images were misclass ied and had metrics less than 1, as sho wn in Figure 4. An instance of the Garuda Pillar , the Dhunge y Dhara, and the Harishankara T emple were misclassied. The performance metrics of misclassied classes are greater than 0 . 96 with the e xception being the recall of Garuda Pillar ( 0 . 9333 ) and precision of V ishw anath T emple ( 0 . 9394 ). The macro a v erage recall ( 0 . 9846 ), precision ( 0 . 9904 ), and F1-score ( 0 . 9872 ) are all abo v e 0 . 98 . The weighted a v erages of performance metrics are also greater than 0 . 98 . This demonstrates that the model performs consistently across the classes. T able 5. Performance metrics of all the classes for the test set Class Precision Recall F1-score Size Bhimsen T emple 1.0000 1.0000 1.0000 42 Char Narayan T emple 1.0000 1.0000 1.0000 35 Dhunge y Dhara 1.0000 0.9615 0.9804 26 Garuda Pillar 1.0000 0.9333 0.9655 15 Harishankara T emple 1.0000 0.9667 0.9831 30 Krishna Mandir 0.9744 1.0000 0.9870 38 Narayan T emple 1.0000 1.0000 1.0000 37 Octagonal De v al 1.0000 1.0000 1.0000 29 V ishw anath T emple 0.9394 1.0000 0.9688 31 Accurac y 0.9894 283 Macro a v erage 0.9904 0.9846 0.9872 283 W eighted a v erage 0.9899 0.9894 0.9894 283 Int J Artif Intell, V ol. 15, No. 4, August 2026: 3326–3341 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3335 Figure 4. Confusion matrix highlighting true and predicted labels between monument classes 4.4. Class-wise err or and r ob ustness analysis In order to understand the model beha vior , a class-wise analysis is performed. In cultural her itage recognition tasks, dif ferent m on um ents present dif ferent le v els of visual comple xity . Thus, e xamining per -class performance helps identify the strengths and limitations of the approach better , compared to an o v erall number . 4.4.1. Class-wise perf ormance v ariability The results indicate high precision and recall across most monument classes. This suggests tha t the model is ef fecti v ely capturing the visual features of monuments. Still, monuments that ha v e clearly identiable structures and unique silhouettes e xhibit higher recognition reliability . The Krishna Mandir has a v ery distinct f acade and structure, and sho ws a near perfect reconition rat e across dif ferent training runs. Y et, a slight performance v ariability is observ ed among a fe w classes. This arises due to se v eral f actors. Man y monuments share similar architectural styles. Harishankara T emple and V ishw anath T emple are almost identical, which e v en confuses the locals of P atan. This is also translated in the model’ s performance, as the model sometimes classies Harishakara T emple incorrectly . This happens when both images are captured from similar vie wpoints. The background clutter and occlusion also af fect recognition. P atan Durbar Square is an acti v e hotspot. Man y of the samples within the dataset also reect this. Because of this, man y monuments are often partially occluded by people or surrounding structures. This mak es it dif cult for models to concentrate on features that are representati v e of monuments. Moreo v er , an edge case in our analysis is when tw o monuments are present within the same image. The Garuda Pillar is surrounded by man y of the other monuments within our dataset. It is also not possible to capture the Garuda Pillar without other monuments, as that doesn’ t depict the real-life scenario. As such, the model sometimes misclassied the Garuda Pillar as the other monument present in the image. 4.4.2. Misclassications patter ns The analysis of misclassied samples sho ws that errors typically occur between monuments that share visual similarity . Some of these ha v e been discussed i n depth in subsection 4.4.1. This highlights that while the model is learning meaningful visual r epresentations of monuments, the y still struggle when discriminati v e A tr ansfer learning-based appr oac h for automatic monument detection (Santosh Giri) Evaluation Warning : The document was created with Spire.PDF for Python.