Research On Small Object Detection Mathod Based On Deep Learning

Small object detection is a hot and challenging research.With the wide application of deep learning in the field of computer vision,object detection algorithms based on deep learning emerge in an endless stream.Compared with traditional object detection algorithms,the performance is greatly improved.And it has been successfully applied in the fields of national defense security,intelligent transportation and industrial automation.Although the emergence of deep learning has promoted the development of the field of object detection,there is still a certain gap between the effect of small object detection and the detection effect of medium and large objects..Intricate scenes,dense distribution of various object,and small object scales all increase the difficulty of feature extraction for small object.This dissertation analyzes and studies the small object detection algorithm based on deep learning.The specific research contents are as follows:First,this dissertation studies and improves the single-scale small object detection algorithm.After researching and comparing different backbone networks,the Swin Transformer with stronger feature extraction ability is selected as the backbone network of the algorithm.And using dilated convolution and residual connection,a more reasonable feature fusion method is designed.The application of this method realizes the fusion of various receptive fields in the feature fusion stage,and makes the network include more features of small object,which effectively improves the performance of small object detection while ensuring the detection efficiency.Secondly,in order to further improve the effect of small object detection,a single-scale high-resolution small object detection method is designed.A feature resolution enhancement module is designed to improve the performance of small object,in order to improve the final feature image resolution for detection.In addition,this dissertation reduces the setting of anchor boxes to make the detection network more friendly to small object.The addition of this module further improves the performance of small object detection.In order to explore the influence of the regression loss function on the network model,various regression loss functions are used to constrain the model.Thus,a more suitable regression loss function is selected.Finally,this dissertation studies and improves the multi-scale small object detection algorithm.The difference between ordinary convolution and involution is explored,and a multi-scale small object detection method based on the fusion of involution and ordinary convolution is designed.The application of involution not only improves the feature extraction capability of small object in the backbone network,but also improves the feature extraction capability of the network for medium and large object.The method continues to integrate the involution into the feature pyramid network,and finally improves the performance indicators of large,medium and small object.