Research On Small Target Detection Based On Deep Learning

In the information age,more and more artificial intelligence technologies are applied in all walks of life.which making people’s production and life easier.As an important direction of artificial intelligence,target detection technology has been widely used in military,transportation and medical fields.However,target detection technology still has some urgent problems to be solved to limit its potential.Low detection accuracy of small targets is one of the problems.For the purpose of solving the problem of low detection accuracy of small targets,this paper explores the reasons for the low detection accuracy of small targets in the Faster R-CNN with FPN structure which is a traditional two-stage detection algorithm.According to the reasons of the low detection accuracy,this paper proposes three improvement strategies,which including using the supervised inception attention network enhances features,designs the initial candidate frame and improves the loss function to facilitate small target detection.Finally,the effectiveness of the improved method is verified through experiments on the MSCOCO2014 dataset.Aiming at the problem that the features of small targets in the feature map are not obvious,this paper proposes to use a supervised inception attention network to achieve target feature enhancement.And theoretically calculate the FPN output used to detect small targets is p2 and p3 channel,then the attention network is applied to the p2 and p3 channels respectively as a feature extraction network.Using the improved method to detect the MSCOCO2014 dataset,the accuracy of small target detection increased by 4.9%.At the same time,the feature extraction network is used as the basic network for subsequent improvements.Aiming at the problem of the large gap between the initial candidate frame and the small target,this paper uses the KMeans algorithm to cluster targets of different scales and design the initial candidate frame.After using this method,the accuracy of small target detection increases by 1.4%.With regard to the problem that the loss function is not conducive to the detection of small targets,this paper improves the DIOU loss function to s_DIOU,and adds scale weight to the loss function to make the network more inclined to predict small targets.After using the improved loss function,the accuracy of small target detection increased by 0.7%.After mixing the two improved methods,the accuracy of small target detection is increased by 1.9%.At the same time,this network is used to detect the other small target dataset which including helmet dataset SHWD and remote sensing dataset Dior.The experiment results also show that this network can effectively improve the effect of small target detection.Above all,it can be shown that the improved methods proposed in this paper has a good improvement on the detection of small targets.