Object detection algorithm based on cosine similarity IoU

Abstract

Aiming at the problem that the traditional IoU-NMS algorithm has poor filtering of redundant boxes with high confidence scores in RetinaNet, an object detection algorithm based on cosine similarity IoU is proposed. Based on the original IoU calculation method, the cosine similarity between the detection boxes is calculated by the vector to better evaluate the similarity between them, which removes redundant boxes with high confidence scores and retain more accurate detection boxes. Meanwhile, CBAM is added to the ResNet-50 network to extract richer semantic information. the detection accuracy of the PASCAL VOC dataset reaches 80.6%, which is 2.1% higher than the benchmark algorithm.