A fast-training GAN for coal–gangue image augmentation based on a few samples.

Data enhancement methods need to be carefully considered and studied for the widespread application of machine vision and deep learning in the mining field. Generative adversarial networks (GANs) prove successful at generating data. However, training a high-resolution image generation network depends on a large-scale dataset and takes a long time. For coal gangue detection, this paper proposes a stride-and-transpose-based progressive generative adversarial network (STP-GAN), which can achieve fast training on a few samples and generate high-resolution images in size of 1024². We employ stride convolutions, up-sampling, and average pooling to construct the model progressively and introduce noise and style optimization. We propose a hidden-layer-frozen progressive training scheme according to the model construction. Compared with other test GANs, STP-GAN generates more authentic and diverse images. The test results of advanced object detection models show that after the auxiliary training of STP-GAN, the mean average precision and average recall of coal–gangue detection are increased by up to 6.92% and 20.39%, respectively. The proposed method can effectively improve the accuracy of coal–gangue detection through data optimization. [ABSTRACT FROM AUTHOR]