IPUN: low-light image enhancement based on image pyramid

In low-light scenarios, the number of photons received by the image sensor decreases, leading to problems such as the lack of color information, poor contrast, and severe noise pollution in the image. Although many effective low-light image enhancement algorithms have been proposed, these algorithms often have difficulty balancing the enhancement effect and operation efficiency. Inspired by U-shaped convolutional neural network (U-Net),this paper proposes Image Pyramid U-Net (IPUN), a lightweight multiscale image enhancement network. By integrating multiscale decomposition with attention mechanisms, the architecture streamlines U-Net while innovatively incorporating feature extraction modules and SCAM attention mechanisms to ensure enhancement efficacy. The design achieves significant performance gains through multiscale feature fusion and attention-guided mechanisms, maintaining computational efficiency. The feature extraction module adopts the idea of traditional image processing and uses the form of convolution to obtain color information, brightness information, and high and low-frequency information in the image. These pieces of information are particularly important for low-light enhancement and can help the network better obtain image features and perform targeted enhancement. Comparative experiments with other algorithms demonstrate the effectiveness and efficiency of our network.