Research And Application Of Image Camouflaged Object Detection Algorithm

When the human visual system observes an image with high contrast between the foreground object and the background environment,it can quickly and accurately search for and locate the target.However,when the foreground object in an image is highly similar to the background or even blends in,the human visual system needs to make several observations to detect the foreground target.To accurately identify foreground targets in images with a high degree of similarity to the background,the task of camouflaged object detection is created.Camouflage is a widespread phenomenon that helps vulnerable creatures escape from natural predators’ complex environments and helps soldiers protect themselves in field practice and combat.Camouflaged object detection aims to detect objects in an image with patterns(texture,density,color,etc.)to their surroundings.Previously,camouflaged object detection relied mainly on hand-crafted camouflaged features,which were complex to produce and less stable and generalizable,resulting in a slow development of camouflaged object detection.To facilitate the development of camouflaged target detection as a task,this paper investigates and designs algorithms for single RGB camouflaged images and groups of RGB camouflaged images,respectively,with the following main work.(1)Detection of camouflaged objects based on intra-layer information enhancement and inter-layer information interactionIn this paper,an intra-layer information enhancement module and a cross-layer information aggregation module are proposed to address the problems of unclear feature representation and inaccurate detection area of camouflaged objects.In particular,the intralayer information enhancement module is used to enhance the channel dimension and spatial dimension features of the information extracted from each layer of the backbone network and uses the element-by-element addition method to transfer information to the features of each layer in the network,and determine the approximate location of the camouflaged target by the enhanced features obtained.Then,the enhanced features of each layer are processed in detail using the inter-layer information fusion module to optimize the target edges while filtering out the underlying background noises to detect the camouflaged target completely model is compared with 12 high-performance target detection models on three publicly available datasets,all achieving optimal detection performance.(2)Collaborative multi-target camouflage object detection for RGB image sets based on feature blending and multi-view exploration In this paper,we construct the first large-scale dataset Co COD8 K,which contains 8528 images covering 5 super-classes and 70 sub-classes,for the multi-target collaborative camouflaged object detection task.In addition,this paper presents the first baseline model for this task-BBNet,which extracts co-camouflaged features and target detail features from within the image group and the image,respectively,and refines the fused information of cocamouflaged features and detail features in both local fields of view and whole global dimensions to map co-camouflaged objects within the image group.The proposed model achieves optimal detection performance compared to 13 high-performance camouflaged object detection models and 6 collaborative object detection models on the proposed standard dataset.