Research On Weakly Supervised Component Instance Segmentation Method Based On Weight Transfer

With the rapid development of intelligent manufacturing,higher requirements are put forward for the accuracy of industrial precision components in the processing and assembly.Visual inspection technology is applied to the inspection and segmentation of industrial precision components due to its non-contact,high accuracy,and good robustness.The weakly supervised instance segmentation methods based on deep learning use weakly supervised labels as pseudo labels to further train the network.However,a large number of industrial component images with bounding box labels are difficult to obtain.At the same time,the uneven illumination of industrial component images and their reflective characteristics will lead to poor quality labels and restrict the improvement of weakly supervised instance segmentation methods’segmentation accuracy.Therefore,studying the method of weakly supervised component instance segmentation,reducing the labeling cost of the algorithm,and improving the accuracy of weakly supervised component instance segmentation has practical significance and value for the subsequent processing and assembly.Based on the component dataset of the obtained bounding box labels,this thesis proposes a component image data enhancement method based on multi-scale homomorphic filtering to solve the problem of insufficient image data of uniformly illuminated components under different illumination characteristics.This method is based on local homomorphic filtering and multi-scale superpixel segmentation is introduced to improve the uneven illumination of component images,and the data is amplified by changing the illumination characteristics of component images to obtain component images with different illumination characteristics under uniform illumination.Combined with component image data enhancement based on geometric transformation and Cycle GAN,enrich the geometric features of component images and transfer their background features,construct characteristic diversity component dataset,reduce high-cost manual annotation,and improve the generalization ability of weakly supervised instance segmentation model.Aiming at the problem of uneven quality of pseudo labels produced by GrabCut due to the smooth and reflective industrial components,this thesis proposes a weakly supervised component instance segmentation method based on weight transfer.The method divides the pseudo labels according to the IoU threshold value between the pseudo labels generated by GrabCut and its bounding box.According to weakly supervised component instance segmentation method based on Mask R-CNN,the weight transfer module is constructed to increase the ability of pseudo labels to learn the relationship between categories,bounding boxes and masks,so as to improve the accuracy of weakly supervised instance segmentation.The experimental results show that using the component dataset test,the mean value and average gradient of the component image data enhancement method based on multi-scale homomorphic filtering are improved by 38.36%and 12.17%compared with the local homomorphic filtering method,which improves uneven illumination of the component images,and it provides component images with uniform illumination under different illumination characteristics.On the component dataset,the proposed method of weakly supervised component instance segmentation based on weight transfer,it’s mAP of Mask and Bbox is 1.6%and 1.5%higher than that of weakly supervised component instance segmentation based on Mask R-CNN,respectively,their mIoU are both improved by 1.8%,which improves the accuracy of weakly supervised component instance segmentation.On the characteristic diversity component dataset after data enhancements,compared with the weakly supervised instance segmentation method based on Mask R-CNN,the proposed method of weakly supervised instance segmentation based on weight transfer,it’s Mask and Bbox mAP are both improved by 2.7%,and their mIoU by 1.8%and 1.9%,respectively,which further expands the range of model accuracy improvement.