Effective low-light image enhancement with multiscale and context learning network

Convolutional Neural Network (CNN) has been widely used in low-light image enhancement task, and has achieved good enhancement results. However, the enhancement results not only are limited by the convolution kernel, but also are affected by the different shapes and sizes of low-light regions. CNN can only capture local dependencies. It is difficult to obtain long-distance dependencies and multiscale features from images, resulting in over/under enhancement. To alleviate these problems, we propose a Multiscale and Context Learning Network (MCLNet) for adaptive low-light enhancement by multiscale feature extraction and global relationships learning. Concretely, in order to obtain discriminative representation in diverse low-light regions, we design an Attentive Residual Multiscale Block (ARMB) to captures valuable multiscale features through spatial attention mechanisms at different scales. Further, we propose a Bottleblock of Scale Aggregation Module (BSAM) to learn hierarchical discriminative features based on th ARBM. Finally, to further adaptive enhancement from globel view, we present a Context Encoding Module (CEM) to model long-distance dependencies by Transformer. Experimental results show that our proposed MCLNet achieves superior performance of low-light images enhancement than some state-of-the-art methods.