Research On Small Object Detection Algirithm Based On YOLOv3

Object detection is a hot and popular research direction in the field of computer vision.It can be found in various fields,such as defense,security and medical security,et al.At present,there are many target detection algorithms,such as Fast R-CNN,Faster R-CNN,YOLOv3.SSD,etc.Where the YOLOv3 algorithm treats target detection as a regression problem on the whole picture,which greatly speeds up the detection,and the backbone network uses a residual structure that greatly deepens the network depth,allowing for comparable detection accuracy.However,these algorithms have a higher rate of missing detection for small targets with an area share of less than 0.12% in the image than for large targets,including YOLOv3.To address this problem,this thesis investigates the small target detection algorithm based on the YOLOv3 network model and works on the following:(1)Since small targets occupy few pixels in the image,the features extracted after multi-layer convolution are not obvious.The improved method enhances the shallow feature information by convolutionally superimposing the 2-fold downsampled feature map from the original network model onto the inputs of the second and third residual blocks,respectively.At the same time,the RFB module is connected after the first 8-fold downsampled feature map to enhance the feature extraction capability.(2)The feature map downsampling operation in the original network with multiple steps of 2 convolutional operations instead of pooled layers reduces the feature transfer capability.To this end,this paper draws on the ideas of Dense Net to transfer shallow feature maps directly to the input of deep homoscale convolutional layers in a densely connected manner.This not only enhances shallow feature reuse,but also effectively mitigates the problem of gradient disappearance.(3)A generalized Io U-based regression loss function is proposed to replace the original regression loss function.The positioning of the predicted box is made more accurate by adding two penalty terms in the loss function: the anchor box is related to the distance from the center of the real box and the predicted box is related to the area of the real box;and at the same time,the problem of gradient disappearance when the two boxes do not intersect is solved.In this thesis,a comparison experiment is performed with the original network on the PASVAL VOC dataset and VEDAI dataset based on the above three improvements,respectively.During training,the training duration was comparable under the same total batch condition.The experimental results show that the above three points have a lower miss rate on small target detection when acting simultaneously on the original YOLOv3 network.The positioning is more accurate and the detection speed is similar.