Research On Weakly Supervised Semantic Segmentation Based On Deep Learning

Image semantic segmentation technique is the key technology of image content analysis and scene semantic understanding and is widely used in medical image analysis,satellite map analysis,automatic driving,video monitoring,and other fields,the full supervision of semantic segmentation model based on deep learning using a large number of data with pixel labeling is training,has obtained the remarkable segmentation accuracy in all fields.However,it takes a lot of manpower and material resources to collect a large number of pixel-level labels,and the final segmentation accuracy of the model is also restricted by the accuracy of data annotation,which makes it difficult to widely apply the fully supervised semantic segmentation model.In combination with the weakly supervised semantic segmentation method,this paper uses imagelevel labels that are easier to obtain to carry out pixel-level segmentation of images and reduce the model’s dependence on pixel-level labels.The main work is as follows:(1)Since image-level labels contain less prior knowledge,such as location information and boundary information,the significance map generated by class-activated mapping only locates part of the significant region of the object,leading to problems such as sparse distribution of objects in pseudo-pixel labels and blurred boundary pixels.This paper studies a multi-scale regional feature and spatial cue context information fusion method.Cascade hollow pyramid network captures the characteristics of the different scale figure to locate the characteristic under the area of objects in different scales,relieve the significant figure in the object localization phenomenon of sparse distribution,characteristics of the context information fusion network level by generating characteristics of pixels affinity matrix to obtain dense space clues context information,using the correlation between pixels characteristics improve the segmentation boundary definition.Experimental results show that the mIoU of the proposed method in the VOC 2012 validation set and test set is 62.1%and 62.4%respectively,which is superior to similar weakly supervised semantic segmentation algorithms and effectively improves the segmentation accuracy of the model.(2)The weakly supervised semantic segmentation method based on class-activated mapping needs to obtain pseudo-pixel labels first,and then use the strongly supervised semantic segmentation method to obtain the final segmentation graph.The training process is complex,and the final segmentation effect is not only dependent on the quality of the generated pseudolabels but also restricted by the precision of the strongly supervised semantic segmentation algorithm itself.In this paper,we study a kind of single phase of the weak supervision of semantic segmentation algorithm,by improving the confidence region partition method for improvement of the confidence region of reliability,the classification of the multi-stage network,and the network embedded in a uniform segmentation model,using the regularization item loss and cross-entropy loss of single-phase model of joint loss training,simplifying the training steps,improve the detection speed and accuracy of segmentation.Experimental results show that the segmentation accuracy mIoU of the proposed method in the VOC 2012 validation set and test set is 63.9%and 65.2%respectively and the reasoning speed of a single image was 29.45ms,which is superior to similar weakly supervised semantic segmentation algorithms,and effectively simplifies the model training process and improves the model segmentation accuracy.(3)Pixel-level annotation of industrial image defect data not only requires a lot of human resources but also has certain requirements on the professional knowledge of the annotators,so it is difficult to obtain a large number of accurate annotations of industrial image defect data.In this article,we will have the bill of lading phase weak supervision and semantic segmentation model migration in the field of industrial image defects detection,industrial image defect data,and natural image data natural gap problem,using the method of image data to enhance and advance training weights migration method to improve the generalization ability of the models,reduce industrial defect image data dependence on pixel-level annotation.Experimental results show that the model segmentation accuracy mIoU is improved by 12%compared with direct migration,and the generalization ability of the model on industrial image data is effectively improved.