关键词: 低照度, 多尺度Retinex, 频域滤波, 模糊集合

Abstract: A new MSR（Multi-Scale Retinex） enhancement algorithm is proposed which is based on frequency-domain to divide scale, in view of halo phenomenon over enhancement and losing the details of the image in the traditional MSR algorithm. The image is converted to frequency domain and three images are obtained via low-pass, band-pass, high-pass filter respectively of the improved algorithm, then the image characteristics are combined with the influence. Finally, in order to reserve image details and solving the problem of halo, the weighted values of the image of three scales are normalized via fuzzy set. The low illumination image is treated as the data source, using PSNR（Peak Signal Noise Ratio）, MSE（Mean Squared Error） and entropy as a quality evaluation index to verify the efficiency of the improved algorithm.