A vision-like network for low-light image enhancement

Zero-shot learning is a popular strategy for low-light image enhancement, as it allows convolutional neural networks (CNNs) to be trained without paired data. However, existing zero-shot learning methods often lead to the loss of image details and noise amplification when processing images with extremely low illumination. To address this issue, a vision-like network is proposed, which is inspired by the human visual mechanism of "first the whole, then the part". This lightweight network has only 5349 parameters and is capable of retaining image details. Additionally, a non-reference loss function based on the median filtering algorithm is proposed to reduce information loss caused by rigid denoising. Experimental results show that our method outperforms existing zero-shot learning methods.