Half-Instance Normalization for Accurate Object Detection in Hazy Conditions

Deep learning models are designed to work in ideal image conditions. Hazy conditions are frequent in drone images. Hence, insufficient lighting and low visibility during image capture significantly degrade the performance of object detection models. Recently, Instance Normalization has been applied in low-level image restoration tasks such as de-raining, de-snowing, dehazing, etc., with significant improvement in the enhancement effect. Inspired by this, a Half-Instance Normalization network (HINet) is proposed as a preprocessing network to assist object detection models by removing hazy regions and enhancing the features of objects. The HINet is trained end-to-end using the A2I2-Haze dataset consisting of hazy and clear image pairs accompanied by bounding box annotations. Additionally, Scaled-YOLOv4 models are trained with a bag of tricks that give further improvements. Extensive experiments have demonstrated the appealing effects of image restoration in increasing the accuracy of object detection tasks.