Research On Camouflaged Object Detection Algorithm Based On Neural Network

With the continuous development of deep learning and computer hardware technology,the hot spot of today’s image research has shifted from research based on traditional methods to computer vision technology based on deep neural networks.Object detection is the most basic problem in the field of computer vision.Its core task is to determine the position and size of a specific target in the image by using a certain target recognition algorithm and search strategy for any given image.As a special target detection task,camouflaged target detection has gradually attracted the attention of scholars at home and abroad in recent years.Camouflaged targets usually have the characteristics of blurred boundaries,confusing coloring and low contrast,which makes it difficult for conventional target detection algorithms to be applied to this type of detection task.The thesis studies the target detection method based on deep neural network,aiming to further explore the lightweight method of the model while ensuring the accuracy of the model.The camouflaged object detection based on deep learning realizes feature learning ability by storing massive weight parameters through multi-layer convolution stacking,pooling and combination.How to design the network to effectively integrate features of different levels and how to effectively remove the background Noise without losing detailed information,and how to achieve generalization capabilities in different complex scenarios are the main challenges in this field.Based on the feature extraction network proposed by the predecessors,the thesis designs an edge extraction method and uses a special feature fusion strategy to make full use of edge information to improve the effectiveness of feature fusion and achieve a fine pixel-level camouflage target detection algorithm.The main work of the thesis is as follows:(1)Aiming at the insufficient edge representation ability of existing camouflaged object detection methods,the thesis proposes a camouflaged object detection network based on graph-guided edge-aware learning.Different from existing methods that use convolutional backbone networks,this method employs a Swin Transformer backbone network to process input images to extract more robust feature representations.Afterwards,a set of edge-aware module(EAM)is designed to realize the prior prediction of the edge of the camouflaged target by using the rich detail information of shallow features.In order to more effectively fuse features of different scales,an edge aggregation module(FAM)is proposed,which uses the graph convolutional network to dynamically learn the inner connection between edge pixels,continuously adjusts the weight of the target boundary,and through the top-to-bottom The next way is to gradually refine the target edge.Extensive experiments were conducted on four mainstream camouflaged target detection datasets,surpassing the existing excellent camouflaged target detection algorithms and achieving the optimal performance.Finally,the effectiveness of different modules is verified by ablation experiments.(2)In order to improve the problem of excessive parameters and high complexity in the above model,and achieve a balance between computational efficiency and prediction accuracy,the thesis proposes a camouflaged target detection network based on edge enhancement and feature fusion.First,replace the more lightweight backbone network to minimize the amount of model parameters;use different inter-layer and intra-layer feature enhancement strategies to make up for the lack of multi-scale representation capabilities of the backbone network;redesign edge extraction module to achieve more accurate edge prediction;finally,the inter-layer attention mechanism is used to achieve the same purification function as the graph convolutional network,which avoids the redundant process of repeatedly sampling edge points,thereby greatly reducing the computational complexity.A large number of experiments have proved that the method achieves a precision superior to other advanced algorithms while greatly reducing the amount of network parameters and computational complexity.Finally,the thesis also conducts experiments on four downstream applications of camouflaged target detection,which proves that the method has a strong generalization ability and has certain application prospects.