Attention-oriented residual block for real-time low-light image enhancement in smart ports

Smart ports utilize an extensive array of surveillance cameras to collect and transmit visual data for timely port monitoring. Nevertheless, complicated and ever-changing lighting conditions, encompassing both natural and artificial sources, frequently lead to low-light images characterized by poor visibility. These kinds of images are insufficient and imprecise in conveying the current condition of the port, posing a risk to essential port operations such as ship identification, containers, trucks, and traffic monitoring. Furthermore, in the edge computing model, the shift of computing resources from the cloud to the edge puts a substantial pressure on the computing capability of edge devices due to the overwhelming influx of visual input from numerous edge sensors. The research proposes an efficient attention-orientated residual block (ARB) for edge intelligence to enhance image quality in low-light conditions while minimizing computation needs. The ARB model consists of three components: a conventional unit (NCU), a channel attention unit (CAU), and a spatial attention unit (SAU). NCU is the backbone in extracting basic features, while CAU employs channel-wise attention to balance the weight of different feature channels. SAU deals with the spatial interactions of significant and less significant elements. These units are connected in parallel to extract crucial information from the feature map while simultaneously increasing the network's ability to convey information in the output map. To further enhance the network's global illumination and local details of low-light images, we embed the ARB model into a multi-scale encoder–decoder framework (termed ARB-Enhancer). We quantitatively evaluated the suggested method's performance by comparing it to current alternative approaches for enhancing synthetic and real-world port datasets. The experimental results demonstrate ARB-Enhancer's superior performance in multiple visual performance evaluation metrics, along with a significant reduction in running time and complexity, making it a promising solution for real-time port monitoring.