Research On Small Target Detection Algorithm Based On Deep Learning

With the rapid development of deep learning and its dominance in target detection tasks,target detection techniques based on deep learning have been widely used in everyone’s life,such as autonomous driving,security monitoring,medical diagnosis,and so on.Nowadays,deep learning target detection has made considerable achievements in the field of conventional target detection,and the detection accuracy of conventional targets has also been greatly improved.However,the shortcomings in the field of small target detection are increasingly evident.At present,there are many problems in small target detection,such as low detection accuracy,prominent false detection and missing detection,which need to be solved urgently.In response to the current situation of insufficient small target detection in the field,this article conducts research on small target detection algorithms based on deep learning.The main work and innovation are reflected in the following aspects:(1)Analyze the difficulties of small target detection and establish a small target dataset.Starting with small target datasets and aiming at the current situation of the lack of small target datasets,a traffic electronic eye dataset(TEE)with minimal targets is established for training and testing small target detection algorithms,and data enhancement techniques are used to solve problems such as data imbalance and image blurring in the dataset.Select the public dataset TT100 K with a certain proportion of small targets as the basis for the algorithm verification work in this article,and then use the Kmeans++algorithm to perform anchor clustering on the training and test dataset.(2)A multi branch attention mechanism module for small target detection is proposed.Aiming at the problem of small target feature information not being obvious and susceptible to background information interference,a multi branch attention mechanism module was constructed by introducing 3-D attention mechanism(Sim AM)and spatial attention mechanism(SAM).Then,a multi branch attention mechanism module is introduced into the YOLOv5 backbone network to enhance the focus of the backbone network on small target feature information,enabling it to extract more small target feature information.Based on the multi branch attention mechanism,a YOLOv5-S detection algorithm is proposed.(3)A multi scale feature enhancement backbone network for small target detection is proposed.Aiming at the problem of insufficient feature extraction ability in current backbone networks when facing small targets,this paper introduces packet convolution to improve the original structure based on the original backbone network CSPMark Net53 of the YOLOv5 algorithm,and combines the multi branch attention mechanism module to build a CSPMarknet-Xt S backbone network.Then,a bidirectional weighted feature fusion network is combined behind the backbone network to increase the fusion and prediction of shallow features,further improving the detection accuracy of small targets.Finally,a YOLOv5-Xt S detection algorithm based on multi-scale feature enhancement of the backbone network is proposed.The experimental results show that the improved algorithm in this paper performs well in multiple small target datasets,with the m AP@50 of the TEE dataset reaching77.6%,the recognition accuracy rate reaching 83.4%,and the recall rate reaching 71.5%.Compared with the previous algorithm,the improved algorithm improves on m AP@50by 12.6%.In the experimental results of the open dataset TT100 K,a high accuracy of m AP@50: 94.3% was also achieved,which has better performance compared to the original algorithm and the latest literature methods.In addition,in the actual detection effect,the algorithm in this paper has a higher detection rate for small targets,and reduces the phenomenon of false detection and missed detection.