多尺度均衡弱光图像增强

Multi-Scale Balanced Low-Light Image Enhancement

摘要: 针对极端弱光图像在亮度恢复、噪声抑制与细节保持方面的矛盾问题，本文提出一种多阶段融合的图像增强方法。在HSV (色调H，饱和度S，亮度V)颜色空间内，通过多尺度高斯滤波估算照度信息，并结合饱和度自适应调节策略，生成两幅亮度候选图，旨在改善图像亮度的不均匀性。非下采样剪切波变换(Non-Subsampled Shearlet Transform, NSST)被用于对亮度候选图进行多尺度分解。其中，低频子带通过主成分分析(Principal Component Analysis, PCA)进行自适应融合以增强全局亮度；高频子带则采用引导滤波和平均融合策略进行处理，以有效抑制噪声并保留结构纹理。融合图像重构后，进一步通过自适应直方图均衡化(Adaptive Histogram Equalization, AHE)优化对比度，最终与原始色度通道结合输出增强结果。实验结果显示，在公开数据集上，峰值信噪比(Peak Signal-to-Noise Ratio, PSNR)和结构相似性(Structural Similarity Index, SSIM)分别达到28.6 dB和0.91，较对比算法分别提升12.3%和8.5%；目标检测的平均精度(mean Average Precision, mAP)达到78.2%。主观视觉评估证实，该方法能有效平衡亮度增强与细节保留，且处理单帧图像仅需0.25秒，具有实际应用价值。

Abstract: In this paper, a multi-stage fusion image enhancement method is proposed to address the contradiction among brightness restoration, noise suppression, and detail preservation in extremely low-light images. In the HSV (Hue, Saturation, Value) color space, multi-scale Gaussian filtering estimates illumination information. Coupled with a saturation adaptive adjustment strategy, this generates two brightness candidate images to improve the non-uniformity of image brightness. Non-Subsampled Shearlet Transform (NSST) performs multi-scale decomposition of these brightness candidate images. The low-frequency subband is adaptively fused via Principal Component Analysis (PCA) for global brightness enhancement. High-frequency subbands undergo processing with self-guided filtering and average fusion strategies, effectively suppressing noise and retaining structural textures. After fusion image reconstruction, Adaptive Histogram Equalization (AHE) optimizes contrast. The final enhanced result is then output by combining with the original chrominance channel. Experimental results on a public dataset show that the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) reach 28.6 dB and 0.91, respectively, which are 12.3% and 8.5% higher than those of comparison algorithms. The mean Average Precision (mAP) for target detection achieves 78.2%. The subjective visual assessment confirms that the method effectively balances brightness enhancement and detail preservation. Processing a single frame image takes only 0.25 seconds, demonstrating practical application value.