Research On Segmentation Algorithm Of Natural Leather Defects In Shoe Industry Based On Convolutional Neural Network

The defect parts on natural leather cannot be used in the production of leather shoes.But at present,the defects can only be detected manually with low efficiency and high labor intensity.Intelligent manufacturing equipment integrated with the function of automatically defect detecting,marking and cutting natural leather is of urgent need.However,the lack of relevant natural leather datasets,the multiple types of defects and the serious interference caused by texture,hinder the detection of natural leather defects in the footwear industry.UP to now,the detection of natural leather defects is generally limited to classification algorithms,and focuses on a very small number of defect types such as tick bites.Therefore,studying the method of natural leather defect detection is the basis of the aforementioned integrated intelligent manufacturing equipment.And is of great significance for improving the production efficiency and the economic benefits of the footwear industry.Firstly,the target detection dataset(Hust?Leather?Obj)of natural leather defect in footwear industry and the semantic segmentation dataset(Hust?Leather?Seg)are constructed.These datasets cover natural leather defect pictures of multiple scale defects,leather colors,textures and defect types.Then a multi-scale target detection algorithm based on adaptive weights is proposed,which effectively reduces the time cost and error of the subsequent defect segmentation process.In the Self-Attention Block(SAB)of the target detection algorithm,the spatial attention and channel attention mechanism are used to make the backbone feature extraction network pay more attention to the defects,and further improve the model understanding ability;while in the Adaptive Feature Fusion Block(AFFB),a more reasonable fusion ratio of deep and shallow features is obtained through trainable weights.Finally,a Boundary-Map Based Algorithm(BBA)is proposed.Among them,the traditional edge recognition algorithm is used to generate key point boundary map in the Boundary Preserve Block(BPB),which is used as an auxiliary input to strengthen the boundary characteristics of the defect;and the consistency score between the prediction map and the boundary map of key points is compared in the Shape Boundary-aware Evaluator(SBE),which is used to enhance the prediction accuracy of the model on the boundary.Experiments on the shoe leather defect dataset Hust?Leather(Obj&Seg)show that the proposed algorithms have a better performance compared with the original benchmark network.In the target detection algorithm,the missed judgment rate MR decreased by8.69%,and the accuracy AP increased by 9.13%.In the defect segmentation algorithm,the intersection(Io U)is improved by 1.33%,and the average pixel accuracy PA is improved by 1.03%.Also,the continuity of the prediction at the defect boundary is significantly improved.Compared with the existing research results,this paper can be applied to more types of defects.