<p>Improving low-light images is challenging due to insufficient lighting, and the task becomes even more difficult when no reference image is available. This limitation has led to the development of zero-shot Low-light image enhancement (LLIE) methods, which do not require paired reference images, making them suitable for real-world applications such as monitoring and autonomous driving. However, many zero-shot LLIE methods aim to adjust illumination, spatial consistency, exposure, and color balancing across all RGB channels. This multi-channel processing increases model complexity, making it harder to handle intricate lighting conditions efficiently. Moreover, models based on standard convolutions have a large number of parameters, resulting in high computational costs. Many methods also fail to prioritize human perception, which is crucial for ensuring perceptual quality and naturalness in enhanced images. To address these challenges, a novel zero-shot LLIE method is proposed that operates on a single channel (the Value channel) in the Hue, Saturation, Value (HSV) color model, simplifying processing and reducing model complexity. This lightweight deep network uses Depthwise Separable Convolution (DSC) to reduce computational costs, with only 4058 parameters, making it suitable for real-time enhancement on resource-constrained devices. The method employs hybrid perceptual losses that combine both computational metrics and human perceptual criteria to guide the enhancement process. It integrates exposure and illumination losses as computational metrics, while the gram matrix texture loss ensures perceptual criteria that aligns with human visual cognition. Extensive experiments validate that our method outperforms existing state-of-the-art approaches.</p>