Research On Image Semantic Segmentation Of Incomplete Label

Fully convolutional neural network has opened up a new development path for image semantic segmentation,which makes image semantic segmentation develop rapidly.Image semantic segmentation is a visual task that resolves different image regions into semantically related categories.Image semantic segmentation has broad research prospects and practical significance in vehicle automatic driving,medical disease diagnosis,and video surveillance.The existing mainstream fully-supervised image semantic segmentation methods require a lot of manpower to perform pixel-level annotation.Therefore,the study of incompletely labeled image semantic segmentation is a practical research direction.However,the existing weakly-supervised semantic segmentation methods that use relatively easy to label samples lack the discussion on the problem of end-to-end weakly-supervised semantic segmentation,the training process is cumbersome and the performance is not good.In addition,the semi-supervised image semantic segmentation methods that use a small number of pixel-level labeled samples,the noise introduced in the consistent regularization method will often lead the network to learn in the wrong direction in the early training period,resulting in a decline in semi-supervised performance.In view of the above content,the research content and main contributions of this thesis include the following three aspects:First,an end-to-end weakly supervised semantic segmentation method based on selfattention mechanism is proposed.Aiming at the disadvantages of the existing weaklysupervised semantic segmentation methods that multiple networks need to be trained separately,the training process is cumbersome and the performance is poor.This method connects the classification network required for weakly-supervised semantic segmentation and the segmentation network using weight sharing backbone network,and proposes a self-attention module based on the class activation map to improve target positioning ability.In addition,a reconstruction branch is introduced to perform image reconstruction tasks,and a self-attention module based on reconstruction features is proposed to refine the segmentation prediction results.This model can not only solve the disadvantages of the cumbersome training process,but also effectively improve the segmentation performance without adding additional saliency processing techniques.The experimental results show that the end-to-end weakly supervised semantic segmentation method designed by this method is simple to train and the model performance is effectively improved.Second,a semi-supervised semantic segmentation method based on consistency regularization is proposed.If only the weakly-supervised semantic segmentation method is used,there is always a big gap between its performance and the fully-supervised image semantic segmentation method.Therefore,appropriate use of a small number of pixel-level samples as training samples can be closer to the performance of the fully-supervised semantic segmentation network.The existing semi-supervised semantic segmentation methods mostly pay attention to the consistency of different disturbances in the image,but the consistency constraint on the error information in the early training period can easily lead to the collapse of the network and the introduction of many irrelevant category prediction problems.This method does not add any noise to the input data,uses the flaw network single-way guidance,and proposes three constraint losses:the single-way guidance constraint between the student network and the teacher network,the segmentation prediction error union constraint,and the feature consistency constraint loss.A large number of experiments show that this method effectively alleviates the consistency constraint of error information in the initial training stage,and effectively improves the performance of semi-supervised semantic segmentation.Third,a semi-supervised semantic segmentation method based on uncertainty is proposed.After using the flaw map in Method 2,it faces the problem of introducing noisy samples.We propose to explain the noise sample’by the uncertainty measurement method of the predicted sample variance,and the uncertainty estimation network is used to evaluate the confidence of the segmentation prediction error.This method designs a semisupervised semantic segmentation method for uncertainty perception,uses the introduced uncertainty estimation network to obtain the high confidence area and the low confidence area of the unlabeled data,and proposes the high confidence segmentation constraint loss and low confidence error constraint loss.The uncertainty measurement of this method solves the problems of the flaw map and accurately evaluates segmentation error.Experimental results show that this method surpasses other latest popular methods and achieves the current best semi-supervised performance.