Channel regularization for lightweight low-light image enhancement

Noteworthy progress has recently been achieved in the development of lightweight techniques for low-light image enhancement. The number of channels in a feature map is pivotal to the performance of a neural network, as it influences the network’s capacity for feature extraction and information representation. However, simply increasing the number of channels by concatenating feature maps can substantially raise computational complexity and introduce a degree of linearity without leveraging nonlinear transformations, thereby limiting the model’s representational capacity. In this paper, we introduce a channel regularization method designed to adaptively evaluate the importance of each channel while maintaining a lightweight architecture to address the computational challenges posed by an excessive number of channels. We propose a Channel Confidence Predictor (CCP) network, which utilizes a lightweight convolutional structure, employing a channel confidence estimator to prioritize channel information with minimal parameter overhead. Additionally, the CCP incorporates a parameter vector for dynamically weighting input features, enhancing the network’s feature extraction and information representation capabilities.