Research On Small Object Detection Algorithm Based On Deep Learning

With the development of science and technology,object detection has become an indispensable part of computer vision.It has high research and application value in many fields,such as face recognition,industrial defect detection,UAV aerial detection,and traffic vehicle detection.Object detection algorithms based on deep learning technology have become increasingly common and mature.However,due to the small proportion of small objects in the image and the few available features,detecting small objects is still a complex problem to be solved.This thesis researches small target detection algorithms based on deep learning to further improve the detection performance of small objects.The main research contents are as follows:(1)To improve the accuracy of the YOLOX-S algorithm when detecting small target objects,an improved algorithm based on multi-scale feature fusion is proposed to improve the detection accuracy and accuracy.The algorithm designs a multi-scale feature fusion network,uses dilated convolution to build a receptive field enhancement module to capture more small target feature information,and uses a large kernel attention method to assign weights to the multi-scale channels of the multi-scale feature fusion network,differentiate according to the importance of input features,automatically ignore the noise response,and increase the attention to small target objects.At the same time,to reduce the amount of model calculation,Res Net50-vd-dcn is designed to replace the original YOLOX-S backbone network CSPDarknet53.In the post-processing stage,the Focal Loss algorithm was employed to rectify the disparity between positive and negative samples,thus augmenting the detection rate of small target objects.(2)A proposal for a parallel multi-branch,high-resolution feature extraction network addresses the issue of low-resolution and the absence of feature data in small target object detection.Construct a parallel multi-depth branch network,with the lower depth network used for high-resolution images and the higher depth network for low-resolution images.Multiple sub-networks containing feature maps of different resolutions are connected in parallel.At the same time,based on parallel connections,continuous fusion is performed between feature maps of different resolutions in the middle position,thoroughly combining high-resolution and lowresolution feature information.(3)Aiming at the problem that the extracted feature information of small objects is limited when the backbone network extracts features from pictures,a feature enhancement method is proposed based on feature map connection flow and attention mechanism flow.The feature map connection flow only uses half the weight of the feature map for the convolutional layer to suppress the number of parameters increased by the convolution.Compared with the general convolution,the amount of learning this way is halved.The attention mechanism flow only performs two simple operations of channel average pooling and sigmoid activation functions without additional learning.The feature enhancement module can enhance the detection accuracy of small target objects without any extra computation.In summary,the method proposed in this thesis is effective.It has been verified by experiments that on the TT100 K dataset,compared with the original YOLOX-S,the improved algorithm has improved the small object detection accuracy by 2.8%,and the small object recall rate has increased by 4.1%.The small object detection accuracy of the proposed parallel multibranch high-resolution network is 2.6% higher than that of Res Net-101 and 2.1% higher than Hourglass-52.On the PASCAL VOC dataset,the m AP of the Retina Net algorithm based on the feature enhancement method has increased by 3.2%.The experimental results show that the method proposed in this thesis positively affects small target detection performance improvement.