Menstrual cycle inspired latent diffusion model for image augmentation in energy production

Abstract

In the energy production domain, image classification is critical for monitoring, diagnostics, and operational optimization tasks. Latent diffusion models (LDMs) have shown potential in generating diverse images during the augmentation process based on text input. However, they are hindered by pixel integrity, texture consistency, and mode collapse. This paper introduces menstrual cycle-inspired latent diffusion model (MCI-LDM), a novel framework that addresses these challenges with three key modifications. First, a menstrual cycle-inspired metaheuristic algorithm is integrated to improve generated images' pixel integrity and structural coherence. Second, an adaptive attention mechanism is employed to dynamically focus on critical regions during image generation, ensuring that fine details are preserved. Third, a multi-scale feature enhancement module is incorporated to capture global structures and local textures, mitigating mode collapse and enhancing overall image quality. Extensive experiments were conducted on five energy-related datasets, demonstrating the superior performance of MCI-LDM in terms of image augmentation, diversity, and generation accuracy. The results highlight the efficiency of the proposed model, making it a valuable tool for improving image classification and data augmentation in energy sector applications. MCI-LDM outperforms LDM by generating more diverse images, with a higher Inception Score (7.1 vs. 5.4) and a lower Fréchet Inception Distance (22.5 vs. 35.2), indicating better quality and variation. Additionally, MCI-LDM preserves image integrity more effectively, achieving superior PSNR (32.7 dB vs. 28.5 dB) and SSIM (0.92 vs. 0.78).