A chemical plant image data enhancement method based on improved SinGAN

Abstract

In response to the limited availability of datasets for abnormal image recognition in chemical plant environments, coupled with challenges such as poor image quality, blurry details and insufficient diversity resulting from traditional data augmentation methods, this paper proposes a novel image data augmentation algorithm based on an enhanced version of SinGAN, specifically tailored to the unique characteristics of chemical plant images. The proposed algorithm improves model generalization by injecting Gaussian white noise into a Gaussian Mixture Model, allowing the complex latent distributions of the noise to more accurately approximate the target data distribution. To further enhance the quality and detail of the generated images, we design a RUBK-Net generator that incorporates a novel Bottleneck Selective Dilated Kernel (BSDK) feature extraction unit as its core architecture. Additionally, we optimize the back-propagation process of SinGAN, significantly reducing training time. Experimental comparisons of nine advanced algorithms, evaluated across four image quality metrics (Fréchet inception distance, peak signal-to-noise ratio, structural similarity and information entropy) on five distinct datasets, show that the improved SinGAN reduces training time by 62.2% compared with the original version, while exhibiting superior image generation performance in chemical plant environments. When applied to image recognition tasks for anomalies, the augmented dataset significantly enhances the generalizability and recognition accuracy of the model.