基于多尺度融合SSD的小目标检测算法

Small Object Detection Algorithm Based on Multi-Scale Fusion SSD

针对一阶段目标检测算法在识别小目标时无法兼顾精度与实时性的问题,提出一种基于多尺度融合单点多盒探测器(SSD)的小目标检测算法。以SSD和DSSD算法的网络结构为基础,设计融合模块以实现Top-Down结构的功能,形成高层网络与低层网络之间的跳跃连接,结合SSD-VGG16扩展卷积特征图以提取多尺度特征,并对不同卷积层、尺度及特征的多元信息进行分类预测与位置回归。在织物瑕疵数据库上的实验结果表明,与SSD、DSSD等算法相比,该算法的检测性能较好,其检测精度达到78.2%,检测速度为51 frame/s,能在保证检测精度的同时提高检测速度。

One-stage detection algorithms cannot balance precision and real-time performance in small object detection.To address the problem,this paper proposes a small object detection algorithm based on multi-scale fusion Single Shot multi-box Detector(SSD).The algorithm designs a fusion module based on the network structure of SSD and Deconvolutional Single Short Detector(DSSD)algorithms to implement functions of Top-Down structure,and thus enables skip connections between high-level network and low-level network.Then SSD-VGG16 is used to extend convolution feature map to extract multi-scale features,and multivariate data of different convolutional layers,scales and features is classified for prediction and position regression.Experimental results on a fabric defect database show that the detection precision of the proposed algorithm reaches 78.2%and detection speed reaches 51 frame/s,which outperforms SSD,DSSD and other algorithms.The results prove that the proposed algorithm can improve the detection speed with a high precision ensured.