The Research And Implementation On Multi-layer Feature Fusion And Optimization In Image Segmentation

The semantic segmentation of images is one of the classic tasks in the field of computer vision.Today,deep learning has been deeply studied by scholars from various countries,and the application of semantic segmentation in many application fields has made breakthrough progress.In the field of image segmentation research,due to the increasingly requirements on the spatial accuracy of segmentation results,research on feature fusion and optimization of segmentation results has been ongoing.In this study,we designed a network that can adapt to a variety of segmentation tasks and have a high degree of feature fusion,and designed an edge optimization method that can be fully integrated into the deep learning network.The comparison of different evaluation indexes with the current methods proves that the model has certain advantages.The main analysis and research contents are as follows.Our paper have made in-depth study on the information that affects the segmentation accuracy,and designed a feature extraction method that fuse representations of different resolution levels.This design can combine the simple features of the low level with the abstract features of the high level more closely,the information contained in the feature is more comprehensive,thus obtaining more accurate results than the ordinary network.The thesis also studies how to design a segmentation network that the network itself can choose the appropriate depth.The model proposed in this paper upsamples feature maps from different depths and combines them with the depth supervision module.By introducing a deep supervision mechanism,the weights of the up-sampled multiple feature maps are distributed.Experiments have achieved good results in the face of different tasks,and the network has a higher adaptability to the segmentation task of different data sets.Based on current edge optimization theories and methods,two options are designed in this study.Most of the optimization is used as a post-processing method to effectively optimize the network segmentations,rather than integrating into the network structure to get an end-to-end model.The scheme designed in this study includes a method of adding edge information by designing a loss function;and a method of optimizing the target category segmentation results by using context semantic information between categories by extracting edges and using the extracted edges as a new classification.Effectively reduce the impact of noise on the results.The experiment used medical data sets and scene segmentation data sets,compared with other models in terms of target segmentation accuracy and model complexity,verified the advantages and disadvantages of the models proposed in the paper and the effectiveness of optimization strategies.