An unsupervised learning method based on U-Net + + for low-light image enhancement

An unsupervised learning method based on U-Net + + for low-light image enhancement

연구 분야: Artificial Intelligence

학회: Signal, Image and Video Processing

In industrial settings like process monitoring and production line optimization, high-quality images play a critical role in managing operational details, rectifying procedural deviations, and improving strategic production outcomes. Images captured in low-light conditions frequently exhibit color distortion, low contrast, and loss of detail, adversely affecting the accuracy and reliability of experimental data. To address these challenges, the study introduces an unsupervised learning method that utilizes the advanced U-Net + + architecture for enhancing low-light images. The advanced U-Net + + architecture, which replaces the traditional U-Net model, enhances deep feature extraction and learning mechanisms, effectively improving image processing for unaligned data. This approach substantially improves low-light photos in brightness, contrast, and color. Enhancements in image quality provide a solid foundation for more reliable and accurate analytical assessments in various applications. Experimental evidence shows that this method not only outperforms existing techniques in visual quality but also excels in key objective metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), contrast difference, and image entropy ratio. This research significantly enhances image quality in low-light conditions using the U-Net + + architecture, improving both visual clarity and informational content crucial for accurate industrial experiments. The advancement promises to boost efficiency in industrial automation, quality inspection, and production management, thereby advancing technological development.