A Lightweight Multi-View Convolutional Neural Network for Cocoon Classification

Cocoon classification plays a critical role in the silk-making process, where accurate sorting of cocoons based on quality is essential for ensuring high-grade textile production. However, due to the variability of cocoon defects and the limitations of single-view image information, achieving high classification accuracy is challenging. This chapter proposes a novel multi-view spatial network to address these challenges. The approach first utilizes a lightweight pre-trained CNN to extract features from each view of a cocoon, including top, bottom, left, and right perspectives. A spatial feature aggregation module is then employed to fuse the information from multiple views, improving classification accuracy by capturing the full range of cocoon features. Furthermore, we designed a multi-view cocoon data acquisition scheme and constructed a large-scale cocoon dataset with expert-labeled data. Experimental results demonstrate that the multi-view spatial network achieves an accuracy of 99.42%, indicating its potential for improving automated cocoon sorting in industrial applications.