Multi-granularity Feature Representation Network Based Camouflaged Object Detection And Its Application

In recent years,with the rapid development of deep learning,the field of computer vision has made great progress.As an emerging direction,camouflaged object detection has attracted the attention and research of many scholars.Camouflaged object detection is designed to detect difficult-to-identify objects hidden in the background,and is now widely used in medical imaging,agricultural detection,military confrontation,industrial quality inspection,work safety and art.Due to the high similarity between the camouflaged object and the image background,it is more challenging than other image segmentation tasks such as salient object detection,and also puts forward higher requirements for feature discrimination.Therefore,in this paper,the feature representation ability of deep convolutional neural network is deeply studied to improve the detection accuracy of camouflage objects and the positioning accuracy of boundary.The main research work is as follows:(1)This paper proposes a camouflaged object detection network model based on multiscale guided correction.Firstly,a multi-scale global sensing module is designed to improve the feature representation capability of the network.In particular,the initial coarse positioning is carried out by cyclic stacking of multi-scale residual blocks on the top of the trunk.Then,a multi-scale guided refinement module is constructed to refine and modify the obtained initial prediction result graph step by step,and the multi-scale information is fused with the multistage side output characteristics.By inserting side output features for multi-scale guidance,the target missing and false detection can be compensated well,and more accurate detection results can be obtained.(2)This paper proposes a boundary guided camouflage object detection network model.Firstly,an initial positioning decoder is proposed to improve the feature representation capability of convolution layer by embedding hierarchical split convolution blocks.After obtaining the initial rough location of the camouflaged target,a boundary-guided doublebranch residual thinning decoder is proposed to gradually repair the missing target part and boundary details.The two decoders are composed of a region branch and a boundary branch for camouflage object detection and boundary detection respectively.In order to make full use of their complementary characteristics,this paper designs an effective boundary-guided region module.With the guidance of boundary features,the region branch is focused on the interior of the boundary for residual learning,so as to achieve more accurate camouflage object detection.(3)A large number of experimental results show that the proposed method performs better than many current advanced methods in four open data sets and in four widely used evaluation indicators.In addition,the network model proposed in this paper is more compact and efficient.In this paper,the proposed method is also applied to agricultural detection,medical imaging and other fields,and good experimental results are obtained.