Region-Specific Adaptive Weighted Multiscale Retinex (RS-AWMSR): A Real-Time Approach to Low-Light Image Enhancement for Sign Language Recognition

Low-light environments have long posed a significant challenge to machine learning models, particularly in real-time applications where accuracy and speed are critical. Under such conditions, many existing methods fail to process images effectively due to insufficient visual information, limiting their practical utility. Sign language recognition is one such use case that is limited by this challenge in image recognition. To address this, a novel methodology—Region-Specific Adaptive Weighted Multiscale Retinex (RS-AWMSR)—is introduced, designed to enhance low-light images in real-time. The approach employs a three-stage process: Preliminary Enhancement, Adaptive Weighted Multiscale Retinex (AW-MSR), and Classification, focusing on the region of interest, specifically the hand region. By targeting these regions, RS-AWMSR not only improves visibility in dark conditions but also maintains computational efficiency, making it suitable for real-time applications. Applied to sign language recognition as a demonstrative case, the method showcases significant improvements in both detection accuracy and image clarity under challenging lighting conditions, offering a robust solution to a longstanding limitation in machine learning-based image recognition systems. The proposed approach holds promise for broader applications in various real-time, low-light image recognition tasks.