Research On The Application Of Small Object Detection Technology Based On Deep Learning Algorithm

Object detection is one of the most basic tasks in the field of computer vision,including two technical problems of classification and location.In recent years,the rapid development of deep learning and convolution neural network technology has pushed the performance of object detection to a new height.However,the detection of small object still faces great challenges.Aiming at the problem of insufficient feature and content representation in small object detection,this thesis proposes optimization schemes from feature scale and content dimension respectively to improve the detection performance of small object.The quality of small object detection depends on the extraction of feature map and the processing of high-level semantic information.For Faster R-CNN,an optimized network of feature scale fusion and enhancement is proposed.The feature scale fusion module realizes the fusion of different scale feature layers by the way of channel number superposition,and the super-pixel convolution adopted does not introduce the calculation amount.The feature scale enhancement module based on multi-scale receptive field and channel attention ensure the robustness of the feature map.The feature map at the top of the optimization network has both the details of the bottom layer and the semantic information of the top layer,which effectively solves the problem of insufficient feature expression of small objects after convolution network.Experiments show that the proposed method has excellent performance for small object detection tasks.The quality of small object detection also depends on the distribution of object content.Firstly,the concept of cleanliness is introduced,and then,GHM Loss is used to improve the loss function.It not only improves the imbalance of positive and negative samples in RPN,but also greatly reduces the huge gap between simple samples and difficult samples in quantity.Then,a small object data enhancement algorithm is proposed,including the data set enhancement strategy based on rotation and the small object instance enhancement strategy based on mask.Not only the data set and the object instance are expanded,but also the training samples are different.Experiments show that the optimized algorithm can greatly improve the detection performance of small objects.