Research On Camouflaged Object Detection Based On Feature Enhancement

Camouflaged Object Detection aims to detect objects hidden in the surrounding environment and can be used in areas such as organ lesion detection and industrial defect detection.Due to the complex environment in which camouflaged objects are located,camouflaged object detection is more challenging.Inspired by the process of human target recognition from simple to complex and from whole to local,the dissertation investigates the method of enhanced feature representation from two stages of overall target prediction and refined overall prediction to alleviate the problem of missing camouflaged targets in ordinary camouflaged scenes,obscured scenes and multi-target scenes,respectively.The main research contents are as follows:(1)To address the problem of the missing edge part of camouflaged object detection,a joint detection method based on neighbor selective aggregation and layer-by-layer adjacent feature enhancement is proposed.The method proposes a layer-by-layer fusion operation of adjacent features to reduce the interference of detailed information on target localization during the generation of initial predictions of camouflaged objects.The layer-by-layer adjacent feature enhancement operation is proposed to enhance the semantic information of the current layer of encoded features using the attention map generated by the previous layer of encoded features to refine the semantic information of the initial predicted features.The comparison experiments,ablation experiments and visualization results demonstrate that the camouflaged object edges detected by this method are clearer and more accurate for overall object localization.(2)To address the problem of insufficient detection integrity in the occlusion environment,a detection method based on layer-by-layer edge and multi-level feature fusion enhancement is proposed.The method ensures the integrity at the occluded edges by enhancing the camouflaged object features with layer-by-layer refined edge features in the decoding stage of the network,and using a contextual aggregation module to fuse multilevel features before the decoder output to mitigate the interference of the occluded object detection by mining the contextual semantics.The comparison experiments,ablation experiments and visualization results demonstrate that the detection method by layer-by-layer refinement of edge information and multi-level feature fusion can better solve the detection of camouflaged objects under occlusion.(3)To address the problem of the missing number of detections in multiple camouflaged object scenarios,an enhancement method based on self-attention for camouflaged object detection is proposed.The method extracts the global information of features and captures the long-range dependencies between objects through the self-attention mechanism in the encoding stage,and compensates for the missing information of feature channels during encoding through dual shared self-attention in the decoding stage,which effectively mitigates the Gaussian fading of the effective receptive field.The comparison experiments,ablation experiments,and visualization results demonstrate that the proposed method can effectively improve the detection accuracy of multiple camouflaged objects,and is also applicable to the problems addressed in the previous two chapters.In conclusion,the problem of missing detection of camouflaged objects from simple to complex scenes can be improved by enhancing the semantic and detailed information of features in the encoding and decoding stages.Experiments on the commonly used public datasets CAMO,CHAMELEON,COD10K,and NC4K confirm the effectiveness of the proposed methods.