Research And Application Of Small Target Detection Based On Deep Learning

Target detection is widely used in precision guidance,medical image diagnosis,automatic driving,image search engine and so on.In the research of target detection,small target detection is always difficult and important.And in practical application,due to the application platform or task,detection can not meet the requirements of accuracy and speed at the same time.Therefore,how to complete the task of target detection quickly and efficiently is of great significance.This paper focuses on the fast and accurent small target detection network and the method of optimizing the on stage detection network.The main contributions are as follows:In order to salve the problems that the feature information of small targets in deep convolutional networks is easily interfered and lost,and ensure the detection algorithm speed,this paper puts forward a small target detection algorithm based on YOLO under the Darknet framework(abbreviated as SOA-Net).To improve the background perception of feature maps,we design a semi-channel feature maps integration structure.Through the combination of the coarse and fine feature fusion,the module merges multi-stage features at a more granular level and increases the number of scales that the output features can represent.Besides,we choose a more robust exponential L1 loss to speed up the elimination of the positioning deviation of the bounding boxes.Then the focal IOU loss is to redistribute the weight of boxes under the guidance of IOU,and increase the weight of boxes with higher IOU among all boxes corresponding to groundtruth.The regression branch gets supervised by the classification loss.The experimental results show that all the optimization parts proposed are helpful to improve the performance of the detection network.The proposed SOA-Net can detect small target objects quickly and precisely.For dealing with the problem of location,even imbalance of positive and negative examples in one stage detection network,this paper analyzes the causes of the problem,and puts forward the optimized training method of target region feature enhancement.In the process of training,we add the auxiliary layer.Ground truth information is introduced to construct eigenvalue matrix,then the input feature map is calculated to enhance the target area according to the targets’ size,so that the network pays more attention to these target areas in the training process,which is helpful to locate targets and restrain the imbalance of positive and negative examples.This strategy is inspired by curriculum learning.With the influence factor gradually reducing,training process becomes harder and harder.This method is only used in the training process,and will not add extra calculation in the actual detection process to affect the detection speed.The experimental results show that the training method proposed in this paper can effectively improve the detection accuracy of the network.