Enhancing the Accuracy of Photofluorescent Uranium Ore Sorting under Dust and Noise Using Swin Transformer and Image Restoration Techniques

摘要: The photofluorescent uranium ore sorting method based on deep learning can effectively separate ore and waste, but its image was easily affected by environmental factors such as dust and noise, resulting in image degradation and low sorting accuracy. Therefore, in this paper, a dust image acquisition system was designed to obtain a test set of synthetic ore images with different dust concentrations, which was used to explore the influence of dust on uranium ore sorting. Five common noises, including Gaussian noise, Rayleigh noise, Gamma noise, Uniform noise and Salt and pepper noise, were used to synthesize ore test sets in different proportions to explore the influence of noise on the accuracy of sorting. The above dust and noise test sets were input into the Swin Transformer model to obtain the accuracy respectively, and the evaluation indicators such as confusion matrix and Grad-CAM algorithm were used for intuitive analysis, and finally, the MIRNet-v2 module was introduced. The results showed that when the dust concentration reached 10g/m³, the accuracy increased from 89.94% to 93.24%. When the noise ratio reaches 0.3, the accuracy of the module for Gamma noise and Salt and pepper noise was increased from 75% and 77.13% to 85.67% and 87.78%, respectively, which has a better removal effect than Gaussian noise, Rayleigh noise and Uniform noise. The research in this paper will provide new ideas for the engineering application of photofluorescent uranium ore sorting methods and solve the influence of complex environmental factors for improving uranium ore sorting.