Object Detection in Rainy Images Based on Multi-stage Image Deraining Network

Automated object detection from a video or image is a tedious computer vision task since it involves localization and object recognition. Most of the advanced driver assistance systems utilize deep learning approaches for detecting the objects. Recognizing objects in rainy environment is a challenging task due to the heavy degradations caused by rain streaks of different sizes, direction, and densities. This leads to the blurring of background scenes and limits the accuracy of object detection models. In this proposed work, a deep learning approach for automated object recognition under rainy scenarios is discussed. This system consists of two modules, deraining module and object detection module. Deraining is performed using a multi-stage architecture based on supervised attention module, and object detection is performed using YOLOv7 algorithm. Single-image deraining is the task of restoring rain-free background image from rain corrupted images. To address the degradations caused by rain, deraining module uses multi-stage architecture which progressively restores the image using supervisory attention signals from ground truth images. This work addresses the challenging image deraining task into multiple subtasks by introducing three different stages along with object detection module. Further this method analyzes the effect of rain in object detection and computes the detection performance in rainy and derainy image situations. Comparative analysis for object detection is done with YOLOv5 and YOLOv7, and the results are quantified. Experimental results with qualitative and quantitative analysis demonstrate that proposed method obtained good performance compared with existing methods.