Overexposed Infrared and Visible Image Fusion Benchmark and Baseline

In night driving scenes, high beams often overexpose visible images, which will cause the failure of many visual intelligence algorithms. The fusion of infrared and visible images can mitigate this problem. However, current infrared and visible image fusion methods are less effective for handling the overexposed challenge. To solve this problem, in this paper, we propose a plug-and-play overexposure prior module which can effectively improve existing infrared and visible fusion methods by integrating it into the network architecture and loss function. Specifically, we first collect a dedicated infrared and visible image fusion dataset OpIVF which focuses on the overexposure scene with 1869 image pairs for training and 40 image pairs for evaluation. Then, we design a simple overexposure prior module based on an adaptive saturation’s overexposure area detection method. The overexposure prior can be regarded as a soft attention map to reweight the contribution of the overexposure area in the source images. Next, to validate the effectiveness of the overexposure prior, we design three kinds of overexposure prior integration methods, including image-based, feature-based, and loss-based. The main goal of these strategies is to retain the information of the infrared image as much as possible in the overexposed areas of the fused image. Finally, we incorporate the overexposure prior into four different infrared and visible image fusion methods. Extensive experiments on the proposed dataset demonstrate that the overexposure prior significantly improves these methods in terms of both the quantitative indicators, visual effects, and downstream visual tasks. The source code and dataset OpIVF are released at https://github.com/xierenping5/OpIVF.