Performance evaluation and model quantization of object detection algorithm for infrared image

Abstract

Object detection technology has always been one of the important research directions in the field of computer vision. It has important application value and prospects in both civil fields and military fields. With the emergence of artificial intelligence technology, deep learning has gradually replaced the traditional algorithm with its higher accuracy. Considering that most of the current algorithms are for color image, compared with color image, infrared image contains less feature information, and it is more difficult in object detection. In this paper, different algorithms are used for object detection in infrared images, and the detection results are compared. This paper chooses YOLO V5 and combines it with MobileNet to lighten the model. After lightening, the parameters are reduced by 30%, but the accuracy is only reduced by 5%. Finally, this paper quantify YOLO V5 based on the model quantization method of PyTorch. After quantization, the accuracy of the model decreases by 2%.