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

In the leather production process, defects on the leather surface are a key factor in the quality of the finished leather. Leather defect detection is an important step in the leather production process, especially for wet blue leather. To improve the efficiency and accuracy of detection, we propose a leather segmentation network using the Kronecker product for multi-path decoding and named KMDNet. The network uses Kronecker products to construct a new semantic information extraction layer named KPCL layer. The KPCL layer is added to the decoding network to form new decoding paths, and these different decoding paths are combined that segment the defective part of the leather image. We collaborate with leather companies to collect relevant leather defect images; use Tensorflow for training, validation, and testing experiments; and compare the detection results with non-machine learning algorithms and semantic segmentation algorithms. The experimental results show that KMDNet has a 1.99% improvement in F1 score compared to UNet for leather and a nearly three times improvement in detection speed.