Vision Transformer-based Deep Learning Model for Leather Species Identification

Identifying the species of animal leather becomes a prominent factor in assessing the quality of the leather goods and also in defending the endangered species. Unlike convolutional neural network (CNN)-based deep learning (DL) architectures, the present work employs vision transformer (ViT)based modelling for leather species identification with a diverse set of images from different regions of the leather. It uses a self-attention mechanism to relate the whole image region enabling the model to analyse the complex pattern efficaciously. An input leather image is divided into disjoint patches and each patch is transformed into vectors. Positional encodings are added to patch embeddings to furnish the spatial hierarchies of each patch inside the image. The resulting vector series are passed through a transformer encoder which aids in capturing the global dependencies. Finally, the species prediction is done by passing the output of a transformer encoder to a multilayer perceptron head. The proposed model achieved a prediction accuracy of 96.3%. Moreover, the experimental analysis compares and proves the superiority of the ViT-based approach over the state-of-the-art CNNs.