Research On Camouflaged Target Detection Algorithm Based On Multi-layer Feature Aggregatio

Camouflaged Object Detection(COD)is a very challenging and important task in the computer field.The COD aims to detect objects hidden in complex environments,and has important practical application value in fields such as medicine,military,and agronomy.Until then,some traditional methods mainly detect camouflaged areas by manually extracting the appearance features of the target,including color,texture,gradient,and other features.Due to the high similarity between the visual features of objects in a camouflaged scene and the limitations of manually extracting features,these methods have poor generalization ability in unpredictable scenes.In contrast,the method based on deep learning has strong flexibility in feature representation,which can effectively solve the above problems.To this end,this paper uses a deep learning-based approach to optimize existing algorithms in terms of multi-layer feature aggregation and model efficiency.The main innovations are as follows:(1)Existing COD algorithms usually ignore the impact of feature expression and fusion methods on the detection performance of multi-level feature aggregation.Therefore,this paper proposes a camouflaged object detection algorithm based on progressive feature enhancement aggregation,and designs an enhancement aggregation network.Firstly,the backbone network is used to extract multi-layer features.Then,in order to optimize feature representation,an enhancement network composed of feature enhancement modules is used to enhance multilayer features;Finally,in an aggregation network,an adjacency aggregation module is designed to fuse adjacent feature information to highlight the features of the camouflaged object area,and a progressive aggregation strategy is proposed to gradually aggregate adjacent features to achieve effective fusion of multi-layer features.Experiments on three common datasets show that the proposed algorithm achieves excellent detection performance.In particular,the enhanced alignment measure and structural measure on the COD10 K dataset reach 0.886 and0.809 respectively,showing the effectiveness of the proposed algorithm.(2)In order to effectively use the semantic information of different layer features and solve the problem that the number of parameters of existing models are usually too large,this paper proposes a camouflaged object detection algorithm based on cross-layer context aggregation,and designs a position refinement network.Firstly,in order to promote the lightweight of the model,Efficient Net-B4 is used to extract multi-layer features.Then,from the perspective of global context information,a special position module is designed for high-level features to achieve effective position of camouflaged objects.Finally,a context refinement module is designed to aggregate cross-level context information,and optimize the position feature map through the detailed information carried by low-level features.In addition,the multi-scale adaptive attention module is used to detect camouflaged objects at different scales during each multi-layer aggregation to improve the detection accuracy and robustness of the network.Experimental results on four public datasets show that the proposed algorithm has excellent detection performance and model efficiency.Especially on the NC4 K dataset,the enhanced alignment measure and mean absolute error reach 0.916 and 0.040,respectively,while the number of parameters and the computation of the model are only 20.64 M and 5.05 G,showing the significant advantages of the proposed algorithm in detection performance and efficiency.In summary,this paper mainly studies the COD algorithm based on multi-layer feature aggregation,among which there are two main works,namely the COD algorithm based on progressive feature enhancement aggregation and the COD algorithm based on cross-layer context aggregation.The experiments of the above two works show that the proposed algorithm has better detection performance,which will help the COD algorithm to achieve further development.