DBLDNet: dual branch low light object detector based on feature localization and multi-scale feature enhancement

In low-light conditions, image details are significantly diminished due to inadequate illumination and low contrast, rendering existing object detection methods are not ideal for low-illumination images. To overcome this challenge, Dual-Branch Low-light Object Detection Network, is proposed, which bases on feature localization and multi-scale feature enhancement. Initially, a low-light image enhancement module is introduced to improve image quality. Adaptive Feature Fusion Module is designed to perform adaptive feature fusion between low-light images and enhanced images, and adaptively adjust their weight ratio, utilizing the detail and texture information from both sources to achieve more accurate object detection results. Additionally, a Feature Localization Module is developed to direct the network focus on the object regions as well as the local features around the object regions, making the object features more discriminative. Furthermore, we designed a Fusion Disruption Module based on dilated convolution with different dilation rates to improve the perception ability of the low-light object detection network to objects of different scales, and DysnakeConv is integrated into the neck of the our network to maintain sensitivity to object edge informations. Finally, the Image Signal Processor algorithm is employed to artificially darken normal-light images, resulting in the VExDark dataset generation, which is developed in conjunction with the ExDark dataset. Comparative experiments with other state-of-the-art low-light object detection methods demonstrate the efficacy of the proposed approach. The dataset has been made publicly available at https://github.com/zzxf123/DBLDNet.