Deep Learning-based Algorithms For Small Target Detection In Video Sequence

With the development of the computer vision technology,increasing object detection technologies are used in autopilot,video surveillance,human-computer interaction,face detection,and other fields.However,owning to the manual designed features of traditional target detection technologies,in which the workload is huge and the robustness of the algorithm is poor they cannot meet the needs of small target detection in the actual video scene.Numerous efforts have thus been devoted to developing new target detection technologies with excellent detection performance.In this sense,the target detection technology based on deep learning may serve as a promising alternative to the traditional technology owning to the superior ability toward extracting image features,which are adapt to various scenes.Therefore,it has important theoretical significance and practical value to study video small target detection technology based on deep learning.A small target detection network based on improved FPN and spatial pyramid pooling is proposed to deal with the problems of small size,low resolution and insignificant features of small targets.We increase the prediction scale of the network by improving the FPN structure of the original network.Meanwhile,the spatial pyramid pooling module in front of the detection layer was also added.The experiment results show that the average detection accuracies of the improved YOLOv3 network are 8.3%,6.1%,and 4.3% higher than those of the original YOLOv3 network model under three scales,respectively.Both the recall rate and accuracy could improve significantly even in the negligible changes of FPS.A small target detection network based on improved dense connection and distributed ranking loss is proposed to deal with the problems of uneven distribution of samples,imbalance between positive and negative samples,and less feature information in the deep layers of small target.First of all,we use Stitcher data enhancement to solve the problem of uneven distribution of small target samples,then propose a new basic network VOVDarknet-53,and finally improve the classification loss function in YOLOv3.The experiment results show that the detection accuracies of the improved YOLOv3 network for small objects and medium objects are 8.3%,6.1%,and 4.3% higher than those of the original YOLOv3 network model.The average detection accuracy(m AP)is increased by 4.1 % even in the negligible changes of the average single image processing time.A model pruning method based on improved channel and layer pruning is proposed to deal with the problem of shortage of computing power and resources in the deployment of small target detection algorithms in real scenes.We improve the channel pruning by setting adaptive local security threshold and carrying out layer pruning by comprehensively evaluating the whole residual structure value,then we apply the proposed model pruning method to detect small targets.The experimental results show that the proposed pruning model achieves higher performance-to-price ratio and is more suitable for small targets detection deployment of small targets in real scenarios.