Fast Deep Multi-patch Progressive Network for Low-light Image Enhancement in Intelligent Transportation Systems

Intelligent transportation systems (ITS) have been widely used in transportation service systems such as roads, railways and waterways. However, the complex low-light imaging environment will lead to the unnatural visual phenomena of over-exposure or under-exposure in the images collected by the visual acquisition equipment of ITS. This paper proposes a fast deep multi-patch progressive network (FDMPN) that aggregates features from multiple image patches in different spatial parts of low-light images with fewer network parameters. Specifically, we propose a multi-receptive field attention module (MRFAM) to enhance local and global feature information perception. To improve the generalization ability of the network, we randomly scramble the smallest feature patches into the MR-FAMs, thereby improving the feature extraction ability of each MRFAM. Experimental results demonstrate that our method can efficiently extract valuable features from complex low-light backgrounds and generate naturally illuminated enhanced images in ITS. Quantitative evaluation results on multiple standard datasets validate that FDMPN outperforms many existing state-of-the-art methods, showing great effectiveness and potential in transportation service systems. The source code is available at https://github.com/LouisYuxuLu/FDMPN.