Research On Object Detection And Segmentation Method Based On Feature Enhancement

Artificial intelligence is currently a relatively popular field.There are many practical application scenarios in life,and the development space is also quite large.Some directions such as detecting and recognizing images and segmenting images have also received a lot of attention.This paper mainly improves the performance of detection and segmentation network through feature enhancement.Specifically,this paper improves the detection and segmentation performance by improving the network structure.In terms of detection,some networks obtain the candidate frame by corner detection,The existing performance bottleneck is mainly corner detection,as the corner is determined by edges,then the accuracy of corner detection can be further improved by adding edge detection branche during training.In terms of semantic segmentation,low-level information is generally guided to the high-level information through skip-connection by introducing deeper output that is stronger in semantic information,that is,high-level information,but there are still some shortcomings.Low-level and high-level information exist domain shift problem,this paper uses domain adaption to convert low-level feature domain to high-level feature domain before feature fusion,and finally achieves the removal of noise caused by low-level detail information problem.In this paper,experiments are conducted on the public data sets COCO,PASCAL VOC,and Cityscapes to verify the effectiveness of feature enhancement in the above two directions.This article delves into the performance enhancement of the better-performing models without the excessive amount of computation.This paper further investigates the problem of instance segmentation by combining detection and segmentation.A simple and effective instance segmentation method with GRU module is proposed,using VoVNet as the backbone network and anchor free method FCOS for object detection,and then performing semantic segmentation on the detection box.The GRU module helps to focus on meaningful features and adaptively forget useless features.The experimental results show that,through such adaptive feature selection,the performance of detection and segmentation can be improved,and the training process is not complicated and is easy to implement.The experimental results show that the feature enhancement method proposed in this paper does not increase the amount of calculation,and the proposed edge detection branch and domain adaptation methods can play a role.Among them,the domain adaptation method has achieved the best results.The proposed feature enhancement methods can be further improved on the basis of strong baseline models.