Abstract: Abstract: Low-light images often suffer from insufficient brightness, severe noise, detail loss, and color distortion, which significantly degrade visual quality and hinder the performance of subsequent vision tasks. To address these issues, a low-light image enhancement network based on lightweight residual and brightness-aware dynamic feature fusion was proposed. Firstly, a lightweight residual feature extraction module was designed in the encoder stage to mitigate information loss caused by downsampling and improve the extraction capability for low-light features. Secondly, a brightness-aware deep semantic processing module was introduced between the encoder and decoder to strengthen the network’s perception and restoration of brightness distribution. Thirdly, a dynamic feature fusion mechanism was applied in the decoder to enhance the integration of skip-connected and upsampled features, thereby improving noise suppression and detail restoration. Furthermore, a perception-color hybrid loss function was proposed to enhance structural consistency and color fidelity of the enhanced images. Finally, a combination of group convolution and Ghost convolution was used to achieve lightweight network design, balancing enhancement quality and computational efficiency. Experimental results on the LOL (LOw-Light) benchmark datasets (LOL-v1, LOL-v2-real, and LOL-v2-syn) demonstrate that the proposed network achieves peak signal-to-noise ratios (PSNR) of 23.71 dB, 21.46 dB, and 24.80 dB, and structural similarity indices (SSIM) of 0.852, 0.863, and 0.933. The overall network adopts a fully convolutional and lightweight architecture. Compared with the lightweight deep curve estimation method Zero-DCE (Zero-Reference Deep Curve Estimation), the proposed method achieves significantly better enhancement quality. Furthermore, compared with EnGAN (Enlighten Generative Adversarial Network), which is based on attention mechanisms and generative adversarial networks, and LLFormer (Low-Light Transformer), which is based on a Transformer architecture, the proposed network substantially reduces model complexity and inference cost while maintaining high enhancement performance. It effectively balances brightness improvement, noise suppression, detail preservation, structural integrity, and color restoration with computational efficiency.

Key words: low-light image enhancement, lightweight residual feature extraction, brightness-aware, dynamic feature fusion, lightweight network design, hybrid loss function

摘要： 摘 要: 针对低光图像常常存在亮度不足、噪声大、细节丢失和颜色失真等问题，提出一种基于轻量残差结构与亮度感知动态特征融合的低光图像增强网络。首先，在编码阶段设计一种基于轻量残差结构的特征提取模块，以缓解下采样过程造成的特征信息丢失问题，提高对低光图像特征的提取能力；其次，在编解码过渡阶段设计一种基于亮度感知的深层语义特征处理模块，以增强网络对低光图像亮度分布的感知和恢复能力；然后，在解码阶段采用轻量级动态特征融合机制，提升跳跃连接特征与上采样特征的融合效果，以提高对低光图像的噪声抑制和细节恢复能力；接着，提出一种基于感知-颜色的混合损失函数，进一步提高低光图像增强的结构一致性与色彩还原度；最后，采用分组卷积与Ghost卷积的组合结构对网络进行轻量化设计，在保证低光图像增强质量的同时降低网络复杂度。实验结果表明，所提出的低光图像增强网络在LOL(LOw-Light)系列数据集(LOL-v1、LOL-v2-real和LOL-v2-syn)上的峰值信噪比(PSNR)分别达到了23.71dB、21.46dB和24.80dB，结构相似性指数(SSIM)分别达到了0.852、0.863和0.933。整体网络采用纯卷积算子架构与轻量化设计，与轻量级深度曲线估计方法Zero-DCE(Zero-Reference Deep Curve Estimation)相比，在低光图像增强质量方面实现了更显著的提升；与基于注意力机制与生成对抗网络的EnGAN(Enlighten Generative Adversarial Network)方法和基于Transformer架构的LLFormer(Low-Light Transformer)方法相比，所提网络在保证低光图像增强质量的同时大幅降低了网络复杂度和推理计算开销，能够在亮度提升、噪声抑制、细节恢复、结构完整性和色彩还原程度等低光图像增强性能与网络计算效率之间实现良好平衡。

关键词: 关键词: 低光图像增强, 轻量残差特征提取, 亮度感知, 动态特征融合, 轻量化设计, 混合损失函数