Small Target Detection Based On Deep Learning

Target detection technology has great application potential in the fields of photoelectric observation tasks,autonomous driving and biomedical.It has played an important role in some scenes in these fields,and it is also the core and foundation of computer vision.Among them,small target detection with a resolution in the range of(0,32] has great application value in both military and civilian fields.The traditional target detection algorithm needs to select the detection window,the manual design feature and then the classification of the classifier according to the specific application scenario.The process is complicated and the model generalization ability is poor.Moreover,small targets usually lack texture features,and the feature dimensions extracted by hand are very low,and the robustness is poor,which can easily cause over-fitting of the classifier.With the development of deep learning,deep convolutional neural networks have achieved incredible results in the field of target detection.A region-based two-stage detection algorithm and a regression-based one-stage detection algorithm are formed,and many excellent algorithms are generated,such as R-CNN,Faster R-CNN,and SSD.However,due to the complex background,occlusion and low resolution,it poses a great challenge to small target detection.Target detection can be divided into two steps: target location and recognition,that is,first find the target location of interest,and then identify the category of the target.After analyzing the classical target detection algorithm based on deep learning,this paper chooses the regionbased two-stage detection idea,and proposes an algorithm for small target detection in the target positioning stage and the target classification stage.The main work and contributions of this paper are as follows:(1)The region of proposal is extracted by RPN network,and the loss function of the predicted bounding box in the RPN network is analyzed.The improved intersection over union IIoU is used as the loss function of the bounding box regression of the RPN network to compensate for the shortcomings of the original Smooth L1 loss function that cannot accurately reflect the overlap between the predicted bounding box and the manually labeled Ground Truth.(2)Each cell of the ROI pooling layer is equally divided into four small regions,and the center point of each region is used as a sampling point,and the pixel values of the four sampling points are calculated by a bilinear interpolation algorithm.The maximum pooling operation is performed on the four sampling points to obtain the output value of the cell..Therefore,the positioning offset caused by the mismatch between the candidate region and the original image region on the feature map caused by the rounding operation in the conventional ROI pooling process is solved.The improved ROI pooling operation has a significant improvement on the small target positioning effect.(3)The feature maps extracted by different depth convolutional networks are analyzed.It is found that the shallow feature maps contain more contours,textures and other information,which is beneficial to small target positioning.The deep feature maps have high-level semantic information,which is beneficial to classification.A multi-convolution feature fusion algorithm based on VGG-16 network is proposed,which makes the merged feature map contain more information to improve the detection accuracy of small targets.(4)The non-maximum algorithm directly sets the scoring of the candidate regions of the same region and the same category to zero,and it can easily miss the overlapping targets.Therefore,the non-maximum suppression algorithm is improved to reduce the missed detection.(5)The algorithm proposed for this small target detection is integrated into a complete algorithm and tested on the TT100 K data set.The test results show that the accuracy and recall rate of the proposed algorithm in the range of(0,32] resolution is higher than that of Faster R-CNN,and the algorithms of Zhu et al.and Yan et al.At the same time,the accuracy of the(0,400] and(32,96] resolution ranges is higher than the above three algorithms.In summary,the algorithm of this paper has a good effect on the detection of small targets in the resolution range of(0,32].