DgrCLR: Dual Gradient Reinforcement Contrastive Learning for Coal Gangue Classification

In the application of coal gangue visual classification, the influence of strong vibration and uneven illumination in an industrial environment causes the fuzzy boundary of the target and reduces the degree of discrimination of sample features. The existing contrastive learning methods cause apparent gradient degradation due to model and design limitations, leading to fuzzy sample classification failure. Getting rid of the samples with poor annotation effects and alleviating gradient degradation are critical issues in improving the accuracy of coal gangue classification. This paper proposes DgrCLR, a dual gradient reinforcement coal gangue classification algorithm based on improved contrastive learning. By designing the Hybrid reinforcement module and Multi-scale feature fusion module, the gradient degradation is alleviated, the dependence on labeled samples is reduced, and the detail representation ability of the model is improved. The experimental results show that the highest accuracy of the DgrCLR model for coal gangue classification is 85.4%, much higher than that of traditional supervised deep learning models. In addition, the gradient strength of DgrCLR is 20 times or higher than that of SimCLR in some model layers by gradient strength summation comparison.