A New Method for Improving the Performance of YOLOv5 in Single Fuzzy Image Recognition

Abstract

Nowadays, multi-target detection technology has achieved high levels of accuracy and has been widely adopted in various domains. However, it is challenging for the multi-object detection model to process fuzzy images and return accurate results in adverse weather conditions such as rainy and foggy. In this paper, a preprocessing method designed specifically for images in rainy and foggy weather is introduced. The reduced visibility in these images arises from the scattering of light by water vapor, which blurs the boundary lines between objects and makes it challenging for the model to accurately identify the object's features. The function of the preprocessing model introduced in this paper is to enhance the distinction between objects in the image, thereby enhancing the object contour and improving the detection rate of the multi-object detection model. The preprocessing model in this paper is calculated based on the third-order matrix composed of image pixels and adjusts the brightness of pixels by calculating the brightness distribution of pixels between rows and columns, so as to realize the increase of the contrast between pixels within a certain threshold range, achieving the feature of strengthening the object boundary while protecting the integrity of the picture. The data and test in this experiment were all based on the model trained by YOLOv5, a representative model for multi-object recognition. The experimental results demonstrated that the preprocessing model introduced in this paper can enhance the recognition ability of YOLOv5 model and improve the recognition rate of YOLOv5.