Research On Camouflage Object Detection Based On Deep Learning

Camouflage is a kind of deception originating in nature.The primary manifestation of camouflage is observed in the evolution of organisms to resemble their surroundings and evade detection by natural predators.Extensive use of camouflage is made in daily life,including soldiers concealing themselves in greenery and donning camouflage uniforms to deceive their adversaries.Detection of camouflaged objects is made difficult due to their high degree of similarity to the surrounding environment.To solve the above problems,two camouflage object detection models based on deep learning are designed.The main research content of this dissertation is summarized as follows:(1)The dynamic interactive refinement network(DIRNet)is proposed for camouflaged object detection,which employs effective details and semantic information to eliminate interference.The bilateral interaction module(BIM)and the adjacent aggregation interaction module(AAIM)are included as components.To improve the accuracy of positioning results for camouflaged objects,interaction-based fusion of foreground and background clues is adopted by the BIM.Ambiguous and incorrect regions in the positioning results are eliminated through the BIM.The AAIM employs an attention mechanism inspired by natural selection to generate cross features.The beneficial features suited for camouflaged object detection are enhanced,while the background noise and useless features are suppressed through the AAIM.Finally,the effectiveness of the proposed model is validated through comparative experiments with other advanced models.And the importance of each module is validated through ablation experiments on each module.(2)The boundary and gradient induced network(BGINet)is proposed for camouflaged object detection,which utilizes boundary and gradient information to improve accuracy in identifying camouflage.The multi-refinement module(MRM),boundary encoder(BE),gradient encoder(GE)and boundary and gradient aware module(BGAM)are included as components.The semantic information is gradually extracted from features to aid detection,using multi-branch multi-scale feature fusion in the MRM.Ambiguity in feature scales and modalities is eliminated through the MRM,allowing for exploration of internal information and the creation of more powerful features.The boundary and gradient encoders are employed to extract corresponding features and produce prediction results for boundaries and gradients,respectively.The extraction of boundary information and gradient information is focused on more by the model through the explicit use of boundary features and gradient features in the BGAM.And more refined prediction results are obtained through the interaction of boundary and gradient,which allows the suppression of background noise.Finally,the effectiveness of the proposed model is validated through comparative experiments with other advanced models.And the importance of each module is validated through ablation experiments on each module.