Local dimming systems with LED backlight arrays improve display quality and reduce power consumption. Traditional parameter-based algorithms often yield suboptimal visual quality, while optimization-based meth-ods are computationally expensive, hindering real-time video applications. This paper proposes a CNN-based Local Dimming Network (LDN) that generates near-optimal backlight luminance matrices by leveraging optimi-zation principles. Compared to parameter-based and CNN-based methods, LDN achieves superior image quality. Unlike optimization-based approaches, LDN maintains high computational efficiency, making it suitable for re-al-time video processing. Thus, LDN balances high visual quality and low computational cost, addressing key limitations of existing local dimming algorithms.