基于改进密集网络与二次回归的小目标检测算法

Small Object Detection Algorithm Based on Improved Dense Network and Quadratic Regression

基于单激发探测器(SSD)的小目标检测算法实时性较差且检测精度较低。为提高小目标检测精度和鲁棒性,提出一种结合改进密集网络和二次回归的小目标检测算法。将SSD算法中骨干网络由VGG16替换为特征提取能力更强且速度更快的DenseNet,利用基于区域候选的检测算法中默认框由粗到细筛选的回归思想设计串级SSD网络结构,在区分目标和背景后进行常规目标分类和位置回归,以获取精确的默认框信息并达到小目标检测中正负样本比例均衡。在此基础上,使用特征图尺度变换方法在不增加参数量情况下完成特征图融合,同时通过K-means聚类方法得到默认框的最佳长宽比并重新设置其尺寸。实验结果表明,该算法的检测平均精度均值在PASCAL VOC2007公共数据集和自制航拍小目标数据集上分别为82.3%和87.6%,较改进前SSD算法分别提升5.1个百分点和9.5个百分点,检测速度达到58 frames/s,可有效实现小目标的实时性检测。

The small object detection algorithms based on Single Shot Detector(SSD)have poor real-time performance and low detection accuracy.In order to improve the accuracy and robustness of small target detection,this paper proposes a small object detection algorithm combining the improved dense network and quadratic regression.The backbone network in the SSD algorithm,VGG16,is replaced with DenseNet that has stronger feature extraction ability and higher speed.The structure of serial SSD network is designed by using the regression idea of coarse-to-fine screening of default box in the detection algorithm based on region candidates.The object and background are distinguished first,and then the conventional object classification and position regression are carried out to obtain accurate default fox information and achieve the balance of the proportion of positive and negative samples in small object detection.On this basis,the feature graph fusion is completed without increasing the number of parameters by using the scale transformation method of feature graph.At the same time,the best aspect ratio of the default box is obtained by using the K-means clustering method,and its size is reset.Experimental results show that the detection mean Average Precision(mAP)of the algorithm is 82.3 percent and 87.6 percent on PASCAL VOC2007 public dataset and self-made aerial photographing small target dataset,respectively,increased by 5.1 and 9.5 percentage points compared with that of the previous SSD algorithm.The detection speed of the algorithm reaches 58 frames/s,which demonstrates that it can effectively realize the real-time detection of small objects.