Research On Object Detection Algorithm Based On Multi-scale Attention And Densely Connected Network

Object detection is a core research direction in the field of computer vision,which aims to locate and classify specific objects in images.With the rapid development of deep learning,object detection is widely applied in many fields such as face recognition,autonomous driving,smart home and security surveillance,which can significantly improve production efficiency,save costs and profoundly affect people’s lifestyles.However,in practical application scenarios,due to the interference of complex backgrounds,insufficient utilization of object feature information and small objects,the object detection accuracy is often not high enough and the missed detection of objects is easily caused.This paper conducts research on the above problems and proposes an object detection algorithm for multi-scale attention and densely connected networks.The main research content is as follows:(1)An object detection algorithm based on multi-scale attention and bidirectional feature fusion is proposed.Firstly,the multi-scale attention module expands the receptive field by multi-branch atrous convolution to obtain multi-scale contextual information,and then coordinate attention(CA)mechanism is used to effectively integrate the spatial position information into the generated feature map,suppressing the background information irrelevant to the object in the image and enhancing the feature representation of the information of interest.Secondly,the deep semantic feature information and the shallow location feature information are effectively fused through a bidirectional feature fusion structure to enhance the feature representation capability at different levels.Finally,the generalized intersection over union loss function is used to solve the bounding box overlap problem and further improve the accuracy of object detection.The experimental results on the Pascal VOC dataset show that the detection accuracy of the model is improved effectively and the missed detection phenomenon of objects is reduced.(2)An object detection algorithm based on densely connected networks is proposed,which uses a densely connected network(DenseNet)to modify the backbone network of the model to reuse features by establishing dense connections between layers of the network,so that every layer obtains feature information from all prior layers and passes feature information from this layer to all following layers,enhancing the propagation of features and further enhancing the utilization of shallow feature information for small objects.In addition,the Mish activation function is used in the bidirectional feature fusion part to eliminate the problem of gradient disappearance in the backpropagation process of the network and to enhance the nonlinear feature extraction capability and generalization performance of the network model,thus obtaining better object detection effect.Experimental results on the Pascal VOC dataset and RSOD dataset show that the model enhances the extraction and utilization of shallow feature information,improves the detection effect of small objects,and also reduces the problem of object miss detection.