A HYBRID APPROACH TO IMAGE ENHANCEMENT IN LOW-LIGHT CONDITIONS BASED ON RETINEX AND RECURRENT CONVOLUTIONAL NETWORKS

Relevance Modern computer vision tasks often face problems with noise, color distortion, and loss of detail, which limits their applicability in real-world scenarios. Existing approaches, including histogram transformations and deep learning-based methods, demonstrate either insufficient adaptability to variable lighting levels or high computational complexity. This creates the need for hybrid solutions that combine the advantages of classical and neural network methods to increase efficiency and versatility. Aim of research The main aim of the research is to investigate a hybrid image enhancement method in low-light conditions that combines the adaptivity of multiscale retinex (MSR-net) with the capabilities of recurrent convolutional neural networks (gruCNN) to integrate time dependencies. The task is to overcome the limitations of the basic methods, such as insufficient processing of dynamic scenes and the inability to maintain the structural integrity of objects when correcting lighting. Research methods To achieve this goal, a hybrid architecture is proposed that combines two key components: multiscale retinex (MSR-net), which provides adaptive illumination correction at different spatial scales, minimizing over-illumination and contrast loss, and the recurrent convolutional network gruCNN, which integrates temporary memory mechanisms (based on GRU blocks) to account for consecutive frames in a video. or related images, which improves resistance to noise and artifacts. The experiments were carried out on LOL datasets (with an emphasis on low illumination) and CIFAR-10 (to assess overall Robustness). SSIM (structural similarity) and PSNR (peak signal-to-noise ratio) were used as metrics, providing a comprehensive quality assessment. Results The hybrid model, combining the methods of Chen Wei-Ting and Zhang Yunhua, surpassed traditional and neural network approaches in improving images in low light conditions. On the LOL dataset, it reached SSIM 0.89 and PSNR 28.6 dB, beating RetinexNet by 12 % and 18 %, respectively. The model effectively restores details, suppresses noise, and preserves the naturalness of images. Thanks to the GRU blocks, it reduced video flickering by 22 %. The model's development prospects will enable image processing in extremely low-light conditions and multispectral imaging. Due to its high accuracy of detail restoration and stability in video processing, the model may be suitable for autonomous driving systems where contextual information is critical, and video surveillance tasks that require reliable operation in real time.