Research On Contextual Information Network And Level Set Based Loss Function In Image Segmentation

In the field of computer vision,image segmentation,as a very key part,is one of the basic problems in the field,and has great value in the practical application.Our research is based on two important sub-fields in image segmentation: semantic segmentation and salient object detection.Semantic segmentation is to understand the image from the pixel level,and it is necessary to determine the corresponding target category of each pixel in the image.The salient object detection comes from visual saliency,that is,to find the target area that the human eye pays most attention to in an image.The salient object detection aims to obtain the salient object,which is a twoclassification problem.Therefore,the salient object detection is biased towards the development of semantic segmentation,which can almost be treated as binary segmentation.A deep research on the network model and loss function has been conducted.On one hand,for the task of semantic segmentation,in terms of network model,a deep learning network model that integrates contextual content is proposed.This model obtains the semantic segmentation of the control target of the correlation between features and features.Experiments prove that the network model in this paper has good segmentation effects.On the other hand,for the task of saliency target detection,in terms of loss function,the energy function thinking of traditional active contour model segmentation methods is introduced to construct a loss function that extracts the adequate features.The energy penalty term is defined so that the proposed loss function can converge quickly.Experiments show that the loss function can converge quickly,can be generalized to many deep learning network models,and can control the process of training to improve the overall accuracy.In the experiment,this paper uses the mean integration of union(m Io U)and mean pixel accuracy(MAP)to qualitatively measure the performance of the method,compares our methods with many traditional models,and also conducts experiments on commonly used datasets.Comprehensive analysis and verification of the applicability and feasibility for our research work well.The contribution of this paper is to solve the lack of image context information in traditional deep neural networks,introduce the relationship between features to build a new neural network.At the same time,it is proposed that combined with the active contour model in the traditional image segmentation field,the format of level set functional is introduced into the loss function,and a new loss function that combines the length and area of the target contour and can converge quickly is consequently constructed.Experiments show that all the innovations in this paper have excellent performance,can be widely used in image segmentation,and have strong segmentation accuracy and generalization ability.