Small Target Detection Algorithm Based On Convolutional Neural Network

Small target detection is a technology that finds intrested objects in different images.This technology has a large number of applications scenarios such as intelligent security,intelligent finance,and automatic driving.Compared with large-scale targets that occupy a larger proportion in the image,small targets are more likely to be ignored by the detector.Improving the feature extraction ability of the small target detection algorithm model and optimizing the parameters of the extracted feature information are benefit to improving the performance of small target detection.With the quickly development of deep learning,target detection algorithms which are based on convolutional neural networks has great advantages in detection accuracy and detection speed.This paper focus on the feature extraction,parameter optimization and sample balance of the small target detection algorithm model.The main research are as follows:(1)In view of the fact that the neural network is more inclined to large targets during the feature extraction process,the feature details that the target detection model can extract are not comprehensive,an improved enhanced receptive field algorithm is proposed.Firstly,a method of increasing neural network receptive field is proposed,so that the network can cover small targets in the process of feature extraction.Secondly,for most target detection network models,the interaction of information is insufficient between different convolutional layers.A method of extracting features detail from different convolutional layers is proposed.Compared to the original features,these features have undergone the scale transformation.Finally,the parameters of vgg16,the main skeleton network the algorithm,are optimized so that the whole model can still remain lightweight.Experiments are carried out on the VOC07+12 dataset,HRRSD aerial photography dataset and MS COCO dataset.The results show that our algorithm has improved the m AP value by 0.5%?18.4%,0.2%?16.4%,1.6%?13.2% respectively on different datasets.(2)In view of the fact that neural networks are easily disturbed by noise information in the process of feature extraction,an improved small target detection algorithm based on weighted network is proposed.Firstly,an improved single channel attention mechanism based on SENet is used to ensure that the network can efficiently obtain effective feature information and suppress interference information during the transmission process.Secondly,in view of the complexity of the sampling weighting process caused by the uncertainty of the sample,a weighting function for the unification of objects is proposed,so that the network can achieve a balance between classification and regression in a multi-task environment.Finally,in view of the small proportion of small targets in the detected image,a data enhancement method is proposed to ensure that the small target has a sufficient proportion in the detected image.Experiments are carried out on the VOC07+12 dataset,HRRSD aerial photography dataset and MS COCO dataset.The results show that our algorithm has improved the m AP value by1.2%?19.1%?1.3%?17.2%?2.1%?13.7% respectively on different datasets.(3)In view of the imbalance between the positive and negative samples in the dataset,which leads to a recall rate of the model,an improved small target detection algorithm based on the equalization of positive and negative samples is proposed.Firstly,it is proposed to rank the positive and negative samples in each training round before screening the positive and negative samples.Secondly,in view of the uneven contribution of the feature information extracted by different convolutional layers during the feature fusion process,it is proposed to give weights that match the contributions of different convolutional layers,and the combination of top-down and bottom-up is used in the process of multi-scale feature fusion,so that feature information of different resolutions can be more efficiently fused.Finally,we apply compound expansion of the network to improve the ability of extracting feature information.Experiments are carried out on the VOC07+12 dataset,HRRSD aerial photography dataset and MS COCO dataset.The results show that our algorithm has improved the m AP value by 3.0%?20.9%?3.4%?19.6%?2.7%?14.3% respectively on different datasets.