用于目标检测的多尺度特征融合算法

In order to solve the general problem that the target detection algorithm is likely to cause missed detection or inaccurate positioning for targets with extreme target size ratios in the data set, this paper starts with the feature extraction algorithm, proposes improvements on the basis of the FPN algorithm, and proposes a set of multi-scale feature fusion algorithms , Combined with the Faster RCNN target detection algorithm, verified on the PASCAL VOC public data set, the experimental results show that the improved multi-scale feature fusion algorithm can effectively extract the features of the target object in the picture. The improved algorithm is compared with the previous one. The algorithm has improved the detection accuracy of most targets, and the overall accuracy of the algorithm has increased by 3.8%. Compared with Cascade RCNN, the most advanced RCNN series algorithm of the target, the improved algorithm in this paper has increased by 0.19%, which shows that this method can effectively improve the extraction of effective features of the target by the feature extraction network.