关键词: 多目标检测, SSD算法优化, 抗旋转卷积层（ARConv）, 概率非极大值抑制（P-NMS）算法, 图片批量测试

Abstract:

As the core of computer vision, target detection is widely used in these aspects, such as face recognition, face tracking, and large-scale scene recognition. In the field of one-stage, the detection speed and detection performance of SSD algorithm are relatively prominent but there are still false detection and missed detection against the complicated background of multi-target detection. Regarding this problem, it proposes multi-target detection method based on improved SSD algorithm to improve testing performance through optimizing network in internal SSD and increasing the applicability of training samples. It adopts modifying network output and optimizing P-NMS algorithm in internal SSD to unify network structure and reduce false detection. It adds ARConv and limiting function to optimize training samples and decrease training time of model. During the period of testing, it proposes single picture batch testing method to effectively increase model recall rate.The experimental results show that improved algorithm has stronger robustness and it can effectively decrease false detection rate to promote network performance.

Key words: multi-target detection, optimization of Single Shot MultiBox Detector（SSD） algorithm, Anti Rotation Convolution（ARConv）, Probability-Non Maximum Suppression（P-NMS） algorithm, picture batch testing