Multi-scale small object detection based on improved Faster R-CNN

Aiming at the problem that the traditional Faster R-CNN is not sensitive to small targets and occluded targets, this paper submits an improved Faster R-CNN target detection algorithm. In this paper, using PASCAL VOC07+2012 to be the experimental data sample set. For the large differences in the targets to be detected in this set, the general anchor size and dimensions is not often used for detecting multi-category problems. For the purpose of increasing small objects detection accuracy, using K-means to improve this situation, the annotation information is centralized for clustering, and the clustering result is replaced by the anchor scale and size in the original RPN. Finally, missed detection and false detection caused by partial overlap of objects in the image, this paper uses the improved soft-NMS algorithm. The experimental results show that, compared with the traditional Faster R-CNN algorithm, the average mean precision (mAP) of the algorithm under the PASCAL VOC07+2012 dataset can reach 80.7%, and it is enhanced by 6.5 percentage points.