SSD small object detection algorithm based on feature enhancement and sample selection
Zhipeng Liu, Wei Hua Fang, Jun Sun · 2021
SSD has a poor detection effect on small objects. The first reasons is its insufficient feature extraction for small objects. To solve this problem, a feature enhancement module is proposed to make better use of the information around the object to improve the identification ability of small objects. The second reason is that the division of positive and negative samples is unreasonable. The threshold is unfriendly for small objects. To solve this problem, an adaptive training sample selection algorithm is adopted to select the threshold. To improve SSD by the above two methods, and experiments on the PASCAL VOC data set. The mean accuracy precision is increased by 2.6% compared to the SSD algorithm. Compared with the series of SSD improved algorithms such as DSOD, RSSD, DSSD, FSSD, the mAP of our method increased by 2.1%, 1.3%, 1.2%, 1.0%. Our method significantly improved the detection effect of small objects, surpassing SSD and its improved algorithms.