Research On Small Object Detection Algorithm Based On Deep Learning In Complex Scene

At present,deep learning-based object detection algorithms such as Faster R-CNN,SSD,and YOLOv5 already have high accuracy and stable detection performance,and their applications are promising.However,there are a large number of small objects in the images captured in application scenarios such as unmanned vehicle autopilot,auxiliary medical diagnosis,traffic accident-prone areas,satellite remote sensing monitoring,etc.Such objects occupy a relatively small portion of the image and do not have obvious features,resulting in unsatisfactory small object detection results.Therefore,the research of this paper focuses on improving the feature extraction capability of small objects to alleviate the above problems.To address the problem of small object information loss caused by the feature extraction process of Faster R-CNN,this paper focuses on the multi-scale feature fusion method and proposes an improved Faster R-CNN small object detection algorithm based on multi-scale auxiliary feature fusion.The algorithm designs a multi-scale auxiliary feature network,which is used to extract shallow features and combine them with the main features,to ensure that even the deepest features have enough spatial information for small target prediction.Meanwhile,the RoI Pooling layer is replaced by the RoI Align layer to make the target localization more accurate.Experimental results show that the detection accuracy of the improved algorithm proposed in this paper is improved by 2.48% compared with the original Faster R-CNN algorithm.To address the problem of insufficient semantic information of shallow features in the SSD algorithm,this paper focuses on feature enhancement and attention mechanism and proposes an improved SSD small target detection algorithm based on the combination of feature enhancement and attention mechanism.The algorithm is based on the network framework of SSD,introducing a shallow feature enhancement network,then using the fusion module to fuse shallow features and deep features,and finally using the channel attention mechanism to fully exploit the small target information.The method can effectively enhance the semantic information of shallow features and reassign weights to different channel features,to effectively extract small target information and suppress background information.The experimental results show that the improved SSD algorithm proposed in this paper has a significant improvement in small object detection accuracy compared with the baseline algorithm.