Target Detection Based On Attention Mechanism And Multi-scale Feature Fusion

The feature pyramid in the existing target detector extracts features from a specific layer of the network for detecting targets,and cannot fully utilize the feature information of feature maps of different scales.And only using local features to detect,can not accurately identify different targets with similar appearance,not suitable for low-resolution image targets,especially the detection of small targets.In response to the above problems,the paper introduces the attention mechanism in the image classification task.From the perspective of enhancing the feature expression ability of the feature pyramid structure,the research combines the global feature information,local information and more efficient multi-scale feature fusion detectors.(1)A target detector combining channel attention and residual learning is proposed.First introduce the channel global attention mechanism,adopt the feature recalibration strategy,learn the weights of different channel features in the feature map through the network,and increase the weight of the effective channel features,thereby enhancing the effective global feature information;then,after the global attention mechanism A lightweight residual block is used as a general function approximator.Through network training residuals,highlight small changes in features,further enhance the expression ability of feature pyramids,and improve the detection performance of small targets in low-resolution images;finally,fuse deep and advanced complex features in shallow feature maps for prediction,improve the detection accuracy of small targets.The algorithm in this paper can accurately detect the target of low resolution image in real time,especially the detection of small target has better effect.(2)A target detector combining attention mechanism and adaptive feature fusion is proposed.First introduce the convolutional attention module to enhance effective global feature information from the channel and spatial dimensions and suppress invalid feature information;then,after the convolutional attention module,use a lightweight residual block as a general function to approximate Device.Training residuals through the network,highlighting small changes in features,combining global and local feature prediction to improve the detection accuracy of similar targets;Finally,using an adaptive feature fusion method,training different scale feature maps through the network The weight of the full use of the feature information of different scale feature maps to improve the overall target detection accuracy.The algorithm in this paper can accurately distinguish background and target in complex scenes,identify different targets with similar appearance,and it is better for small target detection.(3)The algorithm in this paper is validated on the PASCAL VOC 2007,VOC 2012,and MS COCO data sets,and compared with the current mainstream feature pyramid-based target detection method.The experimental results show that the algorithm in this paper improves the performance of low-resolution images,small-size targets,The detection accuracy of multi-scale targets and similar targets between classes,and the target detector combined with channel attention and residual learning,has small parameters and high real-time performance.