Research On Camouflage Object Detection And Military Camouflage Detection Algorithm Based On Deep Learning

As a basic subject in the field of computer vision,object detection is becoming mature and widely used in the daily production and life of human beings,in medical,military,agriculture,mining and art fields,it is found that the detected object can not be well integrated into the background,which poses a great challenge to the existing object detection models and algorithms.Camouflage Object Detection(COD)can better solve the above problems,it aims to hide in the complex background of the target fast and accurate detection.In addition,on the battlefield where camouflage is popular,the identification and location of camouflage objects,especially covert camouflage personnel,is also an important link of modern war reconnaissance.However,compared with salience detection,there are many problems in camouflage detection,such as more complex background environment,more variable shape size,fuzzy contour boundary and more similar color texture,therefore,how to build an efficient and accurate Camouflage Object Detection network is still a challenging task.Inspired by the existing Camouflage Object Detection network,this paper analyzes the advantages and disadvantages of the Camouflage Object Detection network model based on depth learning,and combines a large number of experiments to conduct in-depth research,a deep-shallow feature fusion network with three branches and three levels of supervision is proposed,which is used to detect natural camouflage objects.On the basis of this,combined with the actual military reconnaissance requirements,a detection method of camouflage personnel based on the dual guidance of object edge and texture details is proposed,which is especially suitable for different terrain and environment,camouflage camouflage personnel wearing various types of camouflage uniforms for identification and detection.As follows:(1)When extracting features from backbone network,it is difficult to balance the weight of different kinds of details,which makes it difficult for the algorithm to balance the global and local problems,in this paper,we design a detection model of camouflage object with three branches and three levels of supervision,which decouples the detection task into three parts: object edge,texture detail and image global,and through the three branches of triple supervised learning.In this way,the contradiction between shallow and deep feature extraction can be alleviated by a single branch.Specifically,we use Efficient Net as the context encoder in the main branch to address the more complex background environment and variable target sizes in camouflage detection,the features of the third,fourth and fifth layers are introduced into the trunk branches to extract the spatial features of the deep layers.In order to solve the problem of indistinguishable background,a lightweight edge encoder is designed in the edge-aided branch to extract the edge features of the target,the distance transform method is introduced to transform the original truth map into a new edge detail map,which is used for auxiliary monitoring.In order to solve the problem that the color texture of the camouflage object is similar to the surrounding environment,a texture encoder is designed to extract the texture features of the object in the texture-aided branch,a new texture detail image is obtained by multiplying the gradient and truth maps of the original image element by element.In order to integrate the features of different levels and scales extracted by the three decoders,a feature fusion module is designed.In addition,the traditional binary cross-entropy loss function and cross-parallel ratio loss function are optimized,and a weighted loss function combining the two is designed and applied to the three branches,the loss of the three branches is weighted twice as the final loss function,and the optimal weight parameter is determined by experiment,which optimizes the detection ability of the model to the camouflage object.The experimental results show that the proposed model has good performance on CAMO,CHAMELEON,COD10 K and NC4 K data sets compared with the existing methods.(2)On the basis of the three-branch structure design,we propose a two-stream split detection method based on the dual guidance of object edge and texture details.It is based on Res2Net-50 as the backbone network,still has a trunk branch and two auxiliary branches,the difference is that the two sub-branches do not design a separate encoder but make full use of the different layers of the backbone network to extract image features.The context information extracted from the trunk branch enters the error correction module after the RF module expands the receptive field,the sensitivity degradation caused by network fatigue is mitigated by separating the positive and negative attention flows.The reverse attention flow pays more attention to the boundary information between the camouflage object and the background,we use the edge auxiliary branch to integrate the edge features from the bottom of the backbone into the reverse attention flow through the edge guide module.The forward attention flow pays more attention to the texture details of the camouflage object itself,we use the texture auxiliary branch to integrate the texture details from the middle layer of the backbone network into the forward attention stream through the texture guide module.Different from the previous method,our two auxiliary branches trained the two networks to obtain the edge and texture details respectively in a supervised manner.Finally,we design a cross-level fusion module to improve the final prediction effect.The proposed method is compared with 8 COD methods and one SOD method on the special camouflage personnel data set,and the excellent results are obtained,especially when the boundary of the camouflage object is very fuzzy or the texture is very similar to the background,the effect is obviously better than other existing methods.