Abstract:
Affected by factors such as few light sources and uneven illumination in underground coalmines, underground images have some problems such as low illumination, presenting many dark areas, blurring or missing detail information, excessive darkness generating noise, etc. Traditional image enhancement methods are prone to some shortcomings such as color distortion and loss of detail information in low-light image enhancement. Furthermore, a deep-learning low-light image enhancement method can solve the problem of low-light image brightness enhancement to a certain extent, but its model generalization ability is poor in real-world scenarios. Aiming at above mentioned problems, taking the advantage of Transformer's strong generalization ability, a low illumination image enhancement algorithm based on Transformer model is proposed. Firstly, the Swin v2 module is combined with the convolution module to construct the multiplicative and additive maps of the underground low illuminance image, and superimposed with the original image for fitting, in order to solve the problems of blurring or missing detail information, and over-darkness generating noise. At the same time, the attention mechanism of the fusion multi-scale module is used to perform color processing on the superimposed fitted image to solve the problems of limited image brightness enhancement, the existence of many dark areas, and color distortion. The experiments verify that the performances of this paper's algorithm on the objective quality metrics Peak Signal to Noise Ratio (PSNR), Structural Similarity (SSIM), are improved by 34.76%, 55.73%; 47.32%, 52.76%; 22.52%, 25.7%; 19.615%, 12.285%; 5.81%, 2.625%, compared to LIME, Zero-DCE, RetiNexNet, MBLLEN, and KIND algorithms. Meanwhile, the qualitative analysis results show that the proposed method can significantly enhance the low illumination image of the underground coalmine, the image brightness reaches the visible range, the color is more realistic compared to other methods, and the image detail information is clearer. The study shows that the proposed algorithm has a good performance in terms of the degree of image noise, color distortion, contrast, structural similarity, and brightness, which is relatively superior overall.