Automatic Leather Defect Detection and Classification Using Single-Channel and Multi-Channel Neural Networks

Timely and accurate defect detection is essential in the leather industry, as the quality of raw leather directly impacts both the usability and value of finished products. This paper provides a systematic overview of state-of-the-art solutions and proposes a novel approach for automated detection of leather surface defects using deep neural networks based on the Inception-V3 architecture. Five defect categories are introduced, focusing on their impact on leather quality. In addition, two deep neural network architectures were analyzed and implemented for defect detection and classification: a single-channel model and a multi-channel model with arbitration. The evaluation was carried out using a combination of a custom-developed dataset and publicly available datasets, assessed with standard performance metrics. Moreover, an image annotation tool was developed to facilitate precise defect labeling and the creation of variable-size datasets. Both models demonstrated promising results on the custom dataset, achieving accuracy rates exceeding 93%. The suggested methodology enhances the research domain of leather inspection automation by creating an openly accessible image dataset, performing a comparative analysis of detection models and creating software tools for data preparation. These contributions lay the foundation for further research in leather defect detection and potential industrial implementation.