Lightweight network study of leather defect segmentation with Kronecker product multipath decoding.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface defects significantly impact finished leather quality, necessitating accurate detection methods.
• Wet blue leather defect detection is a critical quality control step in leather production.
• Existing methods lack the efficiency and accuracy required for industrial applications.

Purpose of the Study:

• To introduce KMDNet, a novel leather segmentation network utilizing Kronecker products for multi-path decoding.
• To enhance the efficiency and accuracy of leather surface defect detection.
• To improve the segmentation of defective areas in leather images.

Main Methods:

• Developed KMDNet, incorporating a Kronecker Product Convolutional Layer (KPCL) for semantic information extraction.
• Utilized Kronecker products to create multi-path decoding within the network architecture.
• Trained and validated the model using leather defect images collected in collaboration with industry partners, employing Tensorflow.

Main Results:

• KMDNet demonstrated a 1.99% improvement in F1 score compared to the UNet model for leather defect segmentation.
• The proposed network achieved a nearly threefold increase in detection speed over existing methods.
• Experimental results validated KMDNet's superior performance against non-machine learning and standard semantic segmentation algorithms.

Conclusions:

• KMDNet offers a significant advancement in automated leather defect detection.
• The integration of Kronecker products enhances semantic information extraction and decoding efficiency.
• The proposed method provides a faster and more accurate solution for quality control in the leather industry.