A coal and gangue image classification method integrating DE-DCGAN and RE-EfficientNet

This study proposes a coal and gangue image classification method that integrates DE-DCGAN and RE-EfficientNet to address challenges including inconsistent sample image quality, insufficient morphological coverage in training data, and model deployment difficulties caused by complex underground environments. The method establishes a collaborative framework where an enhanced generative adversarial network (DE-DCGAN) first generates high-quality, semantically consistent synthetic images of coal and gangue by incorporating Exponential Linear Units, dilated convolutional residual blocks, and an efficient channel attention mechanism. Combined with multiple image enhancement techniques, this process constructs an augmented dataset that effectively improves model generalization. Subsequently, the lightweight classification network RE-EfficientNet optimizes deployment efficiency through reconstructed MBConv modules and a streamlined network architecture, maintaining high accuracy while significantly enhancing computational performance. Experimental results demonstrate the method’s effectiveness: DE-DCGAN reduces the Fréchet Inception Distance (FID) by up to 32.70% and improves the accuracy of various classification models by up to 1.88%. Meanwhile, RE-EfficientNet achieves a classification accuracy of 96.00% while reducing parameter count by 42.84%, computational cost by 57.59%, and increasing inference speed by 93.89%. The proposed approach demonstrates potential for practical application and could be beneficial for the intelligent upgrading of coal preparation processes.