ThaiBuddhaPang: An Image Database and Benchmark for Classifying Buddha Images by Pang

Thailand is home to a rich tradition of Buddhist art, and Buddha images serve not only as religious symbols but also as important elements of culture. One primary way these images are identified is by pang, which refers to the specific posture or representation of the Buddha. Different pangs are distinguished by observable visual characteristics such as body posture, hand gestures, and accompanying symbolic elements [1]. Accurate identification of pang can support the documentation and study of digitized Buddhist cultural-heritage collections. As museums, temples, and cultural institutions continue to digitize their collections, computer vision offers useful tools for organizing and analyzing cultural heritage data [2]. Automatic classification can reduce manual annotation effort and improve image retrieval. However, Buddha image classification by pang is challenging. Many classes share very similar overall appearances, and images within the same class may vary due to differences in artistic style, material, age, lighting, viewpoint, background, and image quality, which are characteristics of fine-grained visual classification problems [3]. Deep learning models pre-trained on large-scale datasets like ImageNet have demonstrated strong performance across a wide range of image classification tasks. Architectures such as ResNet [4], EfficientNet [5], MobileNet [6], and Vision Transformers [7] are widely used baselines that can be adapted to specialized domains through transfer learning [8]. Cultural-heritage classification using pre-trained CNNs is an established approach [9]. A further challenge in cultural heritage image analysis is information leakage. When multiple photographs of the same statue or temple appear in both training and test sets, models may partially rely on background or location-specific cues rather than learning the defining characteristics of pang, inflating performance estimates and limiting generalization [10]. However, there has been little published work specifically evaluating these established architectures for Buddha image classification by pang. To the best of our knowledge, no publicly available dataset specifically designed for pang classification has been reported in the literature. This work therefore aims to construct a dataset for pang classification and to establish a benchmark of widely used pre-trained architectures on it. In this study, we initially focus on five pangs: Marawichai, Samathi, Nakprok, Prathanphon, and Saiyat, which are among the most frequently represented postures in Thai Buddhist art [1]. Four architectures are benchmarked—ResNet-18, MobileNetV3-Large, EfficientNet-B0, and ViT-Tiny—trained under a common framework and evaluated using consistent metrics. The central research question of this study is: which pre-trained image classification architecture provides the best generalization performance for Buddha image classification by pang under a statue-instance-disjoint evaluation protocol? This work makes four main contributions. First, it introduces an openly licensed image dataset for Buddha image classification by pang, together with pose labels and statue bounding boxes. Second, it presents one of the first benchmark studies for this task using widely adopted pre-trained deep learning architectures. Third, it introduces a statue-instance-disjoint evaluation protocol built from collection records and a public category graph. Fourth, we quantify the effects of statue-level cropping across four architectures. These findings are intended to serve as a reproducible foundation for future work with expanded datasets, additional pang categories, and more advanced architectures.

The methodology establishes a curated dataset of 826 photographs representing five Buddha pang classes, with related images organized into 77 distinct statue instances using collection metadata and the Wikimedia Commons category graph. Each photograph was assigned a pang label and statue bounding box, allowing the effects of full-image and statue-cropped inputs to be systematically compared. Four established image-classification architectures—ResNet-18, MobileNetV3-Large, EfficientNet-B0, and ViT-Tiny—were initialized with ImageNet-pretrained weights, fine-tuned on the curated dataset, and benchmarked using five-fold statue-instance-disjoint cross-validation.

Images were collected from Wikimedia Commons [11] by crawling categories that correspond to individual Buddha statues, statue families, and iconographic groupings [11]. For each file the crawler records the source category, file name, image dimensions, license, author, source URL, and SHA-1 checksum [11]. Files are de-duplicated on name and checksum at collection time, so no image appears more than once and no image appears under two source categories. Only redistributable images were kept: CC BY-SA (541), CC BY (178), CC0 (69), public domain (37), and one image under the Free Art License. No NonCommercial or NoDerivatives material is included. A source URL is recorded for every image, and photographer attribution for 99% of images. The final dataset has 826 photographs drawn from 92 source categories, covering five pang classes. Table I shows its composition, and Fig. 1 shows two examples of each class.

Two annotators labeled every image through a purpose-built web interface. Each image received one of the five pang labels, or one of two non-target labels. The label other marks a Buddha statue in a pose outside the five target classes, and trash marks an unusable image, with a degree recording whether the cause was viewpoint, quality, or subject. Non-target images are excluded from the benchmark. Each remaining image also received a rectangular bounding box around the statue, drawn in the coordinate space of the stored image. Pose labels and bounding boxes are stored in separate files so that the two annotation tasks could run in parallel.

Photographs of a single statue are near-duplicates of one another. They share the statue, the setting, and often the photographer and the session. A model evaluated on a random split can recognize the statue and recall its class instead of reading the pose [10], so the reported score measures memorization. Our corpus makes that easy to do. The 826 images come from only 77 distinct statue instances. One instance alone accounts for 71 images, and 45 of the 79 Prathanphon images come from a single instance. All images of the same statue instance are assigned to the same fold in cross-validation to avoid data leakage. When source metadata provides category information, images belonging to categories that may contain visually related content are also kept within the same fold. Fig. 2 shows example images from the same category but depicting different statue instances. Under this data leakage prevention protocol, 77 of the 92 source categories in our dataset were identified as corresponding to individual painting or statue instances, including replicas and closely related variants. In cross-validation, folds are formed by stratified grouped sampling, so class proportions are roughly preserved while no statue instance crosses a fold boundary. The code asserts this for every fold. Fold sizes range from 136 to 189 test images, covering 14 to 17 statue instances each. Predictions are pooled out of fold and scored once over the full corpus.

Two input conditions are compared. In the full-image condition the model sees the whole photograph. In the cropped condition the image is first cropped to the annotated bounding box. In both cases the image is then resized so its shorter side is 256 pixels. During training the input is augmented with a random resized crop to 224 × 224 (area scale 0.7 to 1.0, aspect ratio 0.75 to 1.333), followed by color jitter of ± 0.2 in brightness, contrast, and saturation for three of the four models. Such augmentation techniques are commonly used to improve model generalization and reduce overfitting in limited-data image-classification settings [12]. For EfficientNet-B0 the hyperparameter search selected RandAugment [13] (rand-m9-mstd0.5-inc1) in place of color jitter. At evaluation a 224 × 224 center crop is taken with no augmentation. Inputs are normalized with ImageNet channel statistics. We turn horizontal flipping off. Which hand is raised and which way the palm faces are part of what makes a pang what it is, so a mirrored image no longer matches its label.

Four ImageNet-pretrained architectures are compared, listed in Table II: two convolutional networks designed for efficiency, i.e., MobileNetV3-Large [6] and EfficientNet-B0 [5]; one residual convolutional network, i.e., ResNet-18 [4]; and one vision transformer, i.e., ViT-Tiny [7]. All are taken from the timm library. For each model the ImageNet classification head is removed and replaced with a randomly initialized linear layer with five outputs. All parameters are then fine-tuned. Parameter counts are given for the five-class model actually trained, so they are slightly below the published ImageNet figures, which include a 1000-class head.

Hyperparameters were chosen per architecture by an earlier search run under the same statue-instance-disjoint cross-validation protocol, so no setting was picked using the data it is reported on. Learning rate was the only hyperparameter with a consistent effect, and its best value spans a factor of thirty across the four models. Tables III and IV detail the settings. Class weights are set by inverse class frequency, since the largest class is nearly five times the size of the smallest. Weight decay is not applied to normalization and bias parameters, which are one-dimensional. The hyperparameter search picked frozen BatchNorm statistics for EfficientNet-B0. That setting did not train reliably: two runs with identical data, folds, and random seed differed by 0.20 macro F1. BatchNorm statistics are therefore updated during fine-tuning for this architecture. All other settings are as selected. Every arm of a comparison is trained back-to-back inside one process. The pipeline is deterministic given identical inputs, but it is sensitive to small differences in those inputs: reading crops from re-encoded image files instead of cropping in memory shifts macro F1 by up to 0.03 on the same boxes. Control conditions are therefore always re-measured alongside the arm they are compared against, and never taken from stored results.

Out-of-fold predictions are pooled across the five folds and scored once. We report accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1-score, together with per-class precision, recall, and F1, and the full confusion matrix. Macro averaging gives every class equal weight regardless of its size. The class distribution here is heavily imbalanced, so a micro-averaged score would be dominated by Marawichai.

The experimental results show that ResNet-18 achieved the strongest overall performance, reaching a macro F1-score of 0.781 on cropped images. Cropping improved the average macro F1-score by 0.057 across the four architectures, although ViT-Tiny experienced a small decrease. These findings indicate that both model architecture and image pre-processing affect classification performance.

Table V reports macro F1 for all four architectures under both input conditions. ResNet-18 is the strongest on both, reaching 0.781 macro F1 on cropped input, 0.074 ahead of EfficientNet-B0. The ordering is the same in accuracy and in macro precision and recall, as shown in Table VI. Table VI presents the remaining metrics on cropped input.

Cropping to the statue raises macro F1 by 0.057 on average. The gain holds for three of the four backbones and is largest for EfficientNet-B0 at +0.091; ViT-Tiny is the exception, losing 0.011. In a wide temple photograph the statue can occupy a small part of the frame, and cropping puts the hands and posture at a usable scale. The effect is larger than the input sensitivity described in Section II-F, but not by a wide margin, so we report the mean over four backbones rather than any single figure.

Table VII gives per-class precision, recall, and F1 for the strongest configuration. Saiyat is recognized perfectly, with all 146 images correct and no other class assigned to it. Samathi is the weakest at 0.459 recall, and Prathanphon follows at 0.684.

Errors are not spread evenly. Two class pairs account for most of them, and both pairs differ by a small detail rather than by overall shape. Samathi loses 26 of its 98 images to Nakprok and 22 to Marawichai. A naga-sheltered Buddha sits in the same meditation posture with the same hand position, and the serpent hood that separates the two classes sits above the head, which is the part of the frame a tight crop is most likely to lose, as Fig. 3 illustrates. Most of the damage comes from one statue instance: the Emerald Buddha and its replicas contribute 46 Samathi images, of which only 5 are classified correctly. Every other Samathi instance reaches 0.67 recall or better. Prathanphon splits into two presentations that the label does not distinguish. Taking box aspect ratio as the divider, the 59 standing images reach 0.881 recall, while the 20 seated images reach 0.100, and 15 of those 20 are assigned to Marawichai. A seated boon-giving Buddha holds one hand at the knee in the same place as an earth-touching Buddha, and the two differ only in whether the palm faces outward or the fingers point down.

ResNet-18 is the largest of the four models by parameter count and also the strongest, but the ordering does not follow size: EfficientNet-B0 and MobileNetV3-Large are the two smallest and finish second and third, while ViT-Tiny is last by a clear margin. With 826 training images the transformer has too little data to learn the attention patterns it depends on, whereas the convolutional models arrive with a spatial prior already built in. Cropping helps more than the choice of architecture for three of the four models. The gain from cropping averages 0.057, and the gap between the best and second-best architecture is 0.074, so the two effects are of comparable size. Anyone extending this benchmark should treat the input pipeline as a first-class variable rather than fixing it and tuning only the model. Statue diversity is the ceiling. The 826 images rest on 77 statue instances, and the two smallest classes have fewer than ten instances each. Samathi draws 46 of its 98 images from the Emerald Buddha and its replicas, and those 46 supply 5 correct predictions between them. More photographs of statues already in the set cannot fix that. Only photographs of statues that are not yet represented will.

The dataset draws mostly from Thailand but also includes material from India, Nepal, Myanmar, Cambodia, Sri Lanka, Tibet, Korea, and Bhutan. Two classes rely on it heavily: 84% of Prathanphon images and 64% of Nakprok images come from outside Thailand or from categories that do not record a country, because Thai examples of those poses are scarce in the source repository. Replicas are the weak point of the protocol. If the category hierarchy does not record that one statue is a copy of another, we do not catch the link, and the two can land in different folds. Four such relationships are documented in art-historical sources but absent from the hierarchy; adding them by hand changed macro F1 by less than the run-to-run variation of the pipeline.

We introduced an openly licensed dataset of 826 photographs of Buddha statues covering five pang categories, with pose labels and statue bounding boxes, and benchmarked four pre-trained architectures on it. ResNet-18 performs best at 0.781 macro F1 on cropped input, and cropping to the statue is worth 0.057 macro F1 on average. All results use five-fold cross-validation over statue instances identified automatically from collection records and a public category graph, so the protocol can be reproduced and extended without curating the objects by hand. The error analysis points at the same place twice. Samathi against Nakprok and Prathanphon against Marawichai both turn on a small region: a hood above the head in one case, the rotation of a palm in the other. Whole-image classification has no way to attend to those regions specifically. Locating the hands and the area above the head before classification is the most promising next step, and the dataset released here provides the boxes needed to start.

The authors gratefully acknowledge the Sirindhorn International Institute of Technology (SIIT), Thammasat University, and Queen Mary University of London (QMUL) for their financial support. This support was granted through the International Science Partnerships Fund (ISPF), administered by the British Council, to promote academic collaboration between QMUL and SIIT.