关键词: 小目标检测, 特征融合, SSD（Single Shot Multibox Detector）, 特征增强, PASCAL VOC2007

Abstract:

Aiming at the shortcoming that SSD algorithm has poor detection effect on small targets, an improved SSD method is proposed based on feature fusion. The SSD method of feature fusion is proposed. This method fully integrates the deep and shallow feature information to improve the network model’s ability to detect small targets. In order to better detect small targets, this paper adjusts the prior frame size relative to the original image ratio column. At the same time, the corresponding hyperparameter values of the SSD model are adjusted. The experimental results show that the detection accuracy mAP is 3.4 percentage points higher than that of SSD, and the detection accuracy of small target Bottle, Chair and Plant is increased by 8.7, 3.4 and 7.1 percentage points, respectively. The detection accuracy mAP is significantly improved compared to the state-of-the-art target detection algorithms. In addition, the extended experiments show that the proposed algorithm can successfully detect the small targets that were not detected by traditional SSD, and improve the average detection accuracy.

Key words: small target detection, feature fusion, SSD（Single Shot?Multibox Detector）, feature enhancement, PASCAL VOC2007