Research And Implement Of Image Segmentation Algorithm Based On Object Extraction And High-Resolution Network

How to fine-grained recognize the content in the image is the focus of further image processing.Accurate image segmentation algorithm is the key way to solve this problem.Earlier image segmentation techniques are limited by low-level pixel features,overly rely on certain mathematical assumptions,and lack of the ability to extract high-level semantic features.Existing image segmentation technologies utilize deep convolutional networks for end-to-end image segmentation,which are able to extract semantic information.However,there are problems such as poor generalization performance,inaccurate segmentation edges,and low segmentation accuracy.Therefore,segmentation results are quite different from the expected.Accurate image segmentation needs to solve several key issues:first,overfitting and poor generalization performance due to scarcity of datasets;then,how to make full use of data sample similarity to improve segmentation accuracy;and finally,how to eliminate isolated regions in segmentation results and optimize segmentation edges.In response to the above-mentioned key scientific issues,the main work and innovations completed in this thesis are as follows:(1)Aiming at the problem of reducing the cost of dataset annotation,a semantic segmentation-based image object extraction algorithm(SSOE)is proposed.The algorithm converts the prior information into a tensor form that can be processed by the neural network through the fast preprocessing of the binary plane;accurately extracts the semantic information of the target area through the improved semantic segmentation model,and predicts the mask of the target area;through the auxiliary loss function Label Loss speeds up training convergence.Experimental results show that the algorithm has the advantages of high accuracy,fast inference speed and short training time.(2)Aiming at the problem of exploiting the similarity of data samples,a difference prediction algorithm based on similar samples and highresolution networks is proposed.The algorithm selects different data samples through the image similarity algorithm;uses the encoder-decoderbased difference prediction network to predict the segmentation mask differences of the data samples;uses the multi-reference sample voting strategy to eliminate accidental errors and optimize the segmentation results.Experimental results show that the algorithm can achieve the current highest segmentation accuracy on multiple datasets.(3)Aiming at the problem of optimizing segmentation edges,a postprocessing algorithm based on connected prior is proposed for the problem of.The algorithm filters out the isolated regions through the connected region screening strategy based on the contour labeling method,and optimizes segmentation edges;isolated regions are re-segmented to the correct region through the isolated region re-segmentation algorithm based on the probability map.The experimental results show that the algorithm can significantly repair the wrong segmentation edges,so that the accuracy of the segmentation results is improved over other image segmentation post-processing algorithms.