Research On Semantic Segmentation-Oriented Attention Mechanism And Multi-Scale Feature Cross-Layer Fusion

In recent years,the development of deep learning has been changing with each passing day.With the continuous efforts of scientists and scholars arou nd the world,there are more and more deep learning projects that can be applied to all aspects of our lives in our real life.The semantic segmentation studied in this paper is actually a intensive classification task,which is equivalent to classifying each pixel in the picture,and finally masking it to represent the category of a specific set of pixels on a macro level.Nowadays,semantic segmentation has been widely used in the fields of unmanned driving,remote sensing images,medical images and so on.However,with a large number of applications of semantic segmentation,different problems have also appeared in practical applications.For example,the network cannot learn objects with complex features well,resulting in poor segmentation of complex objects and the network's insensitivity to multi-scale information.This leads to poor segmentation of objects of different sizes.This article believes that solving these two problems is of great significance for promoting the implementation of semantic segmentation in actual scenes.First,this article starts with the attention mechanism,and improves the network's ability to learn objects with complex features by designing an attention module that meets the needs of the scene.Secondly,the multi-scale feature fusion structure is used to enhance the network's ability to recognize objects of different scales.Therefore,in response to the above problems,the work carried out mainly includes the following two points:(1)In order to improve the sensitivity of the network to the features of complex objects,this paper combines the attention mechanism in different scenarios and proposes the CAM module.This module tries to use high-level features containing abstract semantic information as a guide to the underlying features containing spatial detail information are screened to achieve the purpose of suppressing invalid spatial detail features and emphasizing effective spatial detail features.At a macro level,this is con ducive to the network's recognition of objects with complex features.(2)In order to enhance the network's segmentation ability on objects with different segmentation scales,this article starts from the aspect of cross-layer fusion of multi-scale information,and studies how to fuse information of different scales will cause problems in identifying different scales to meet expectations.effect.And continue ablation experiments for different network structures.In this paper,different experiments were done on the famous street view dataset cityspaces,a dataset of esophageal tumors representing medical images.The experimental results show that the channel attention mechanism can indeed improve the segmentation effect of the network very well in the esophageal tumor data set with complex characteristics,and the combination of multi-layer fusion of modules of different sizes can further improve the performance.In the street view dataset cityspaces,the experiment also proved the effectiveness of the multi-scale fusion feature network.