Research On Target Detection Methods In Complex Environments

In the current era,the increasing progress of deep learning technology has opened a new path for research and development in the field of computer vision.As a key research direction in the field of computer vision,object detection has attracted more and more researchers' attention.In recent years,the object detection approaches based on deep learning have significantly replaced the object detection approaches based on handcraft features and machine learning,and have made breakthrough progress in detection accuracy and speed.In most current object detection methods,in order to improve the detection accuracy of the model and speed up the detection speed of the model,large and medium object with rich feature information have obtained better detection results.In contrast,the detection effect of small object needs to be improved.At the same time,due to the complexity of the actual environment,the object will present different forms in different environments,and there will be various complex occlusion problems that need to be solved.For this reason,based on the existing research methods,this article focuses on the target detection algorithms in complex environments,mainly including the following three aspects:First of all,this paper takes the attention mechanism as the starting point,analyzes the existing attention mechanism-based algorithms,and proposes a Non-local Channel Attention Block(NLCA Block)based on its characteristics of the focus area.This Block adaptively learns the attention weights simultaneously in the spatial domain and channel domain of the feature map,acquires the information that needs to be focused on in the feature map,and suppresses redundant background information,thereby enhancing the ability to express target detailed features.Experiments on image classification and object detection tasks and the results show that the NLCA Block can effectively improve the performance of the basic network,and has universality,which is suitable for various basic convolutional neural network frameworks.Secondly,considering the small proportion of small object in the image,as the number of layers in the target detection network deepens,the problem of lack of feature information of small object will arise.This paper proposes a new small object detection network,which uses the attention mechanism in the shallow layer to improve the ability to distinguish features and enhance the context and semantic information of small object in shallow features.At the same time,the feature extraction capability of deep networks is enhanced through densely connected structures to obtain rich object information so that the object takes into account both semantic and detailed information.The experimental results show that the algorithm has better detection results in the PASCAL VOC and MS COCO data sets,and can effectively improve the detection accuracy of small object targets on the premise of ensuring the detection speed.Finally,based on the analysis of previous target detection methods,it is concluded that the receptive field is of critical significance for object detection at different scales.Aiming at the problem of occlusion between targets in complex environments,this paper proposes a Multi-scale Receptive Field Feature Fusion Network(MR3FN).The network is designed with a multi-branch receptive field module.Multi-scale receptive fields are realized by using extended convolution to obtain object context information of different receptive fields,which enhances the diversity of object features in the network.In addition,through the feature fusion module,object features with different expression capabilities are fused to obtain strong semantic information,thereby improving the expression capability of the entire network feature.Repulsion Loss is further used to optimize the object detection model to enhance the detection performance of the model in the case of object occlusion.The experiment proves that the MR3 FN has achieved a good detection accuracy rate in the UA-DETRAC dataset and effectively improved the object detection effect in the case of occlusion.