Research On Small Object Detection Method Based On Deep Learning

Object detection is a fundamental task in the field of computer vision,laying the foundation for the development of other directions in computer vision.Its main purpose is to find the location of the object from the image and identify the corresponding category of the object.Small object detection,as an important branch of object detection,is widely used in fields such as intelligent driving,urban road traffic,and industrial automation detection.With the development of deep learning,object detection algorithms based on deep learning have achieved significant results.However,the performance of detecting small and medium-sized objects in images is still not ideal compared to medium and large-scale objects.Small objects have the characteristics of fewer pixels,denser distribution,and higher positioning accuracy requirements for regression boxes,making small object detection more difficult.This article analyzes and studies the problems in the detection of small objects based on deep learning.The specific work is as follows:In order to solve the problem of poor accuracy in small object detection,a graph neural network-based decoupled small object detection network is proposed from the perspective of containing fewer small object features in the feature map.Firstly,use the CSPPark Net-53 backbone network improved by YOLOv5 to extract features.Secondly,by constructing a graph neural network to calculate attention for each node in the graph,mapping back to the feature map to compensate for the lost detail information in the feature fusion process of the feature pyramid network.Once again,the decoupling head structure is used to decouple the classification and regression tasks.Finally,a dynamic top-K strategy is used to match more suitable prediction boxes for each GT(Ground Truth)box.To solve the problem of difficult and easily missed detection of small objects in dense scenes,a Transformer based small object detection method is proposed,which considers object detection as a direct set prediction problem.Firstly,the feature extraction network was improved.Secondly,incorporating position encoding into the encoder decoder structure of Transformer effectively improves the positioning accuracy of the prediction box.Once again,compensate for decoder defects caused by content queries through Mixed Query Selection.Finally,the GT box is matched with the prediction box through a one-on-one matching strategy.In order to apply the model to practical engineering,a semi-supervised distillation model for small object detection was proposed.Firstly,the teacher student model is used to update each other’s weights to train the model.Secondly,a lightweight single scale object detection network is used to extract high-resolution single scale feature maps through Swin-L.Thirdly,the Receptive field of the single scale feature map is expanded by expanding convolution.Finally,the Box Jitting strategy and balanced sample matching strategy were adopted for small objects to improve the model’s detection performance for small objects,and good detection results were obtained on public datasets.