基于投影寻踪子波学习网络的图像无监督恢复

On Improving Unsupervised Restoration of Image with PPWLN

针对现有模糊图像的复原方法 ,提出了一类新型人工神经网络——投影寻踪子波学习网络 ,并将其用来处理图像的去模糊问题。这类新型网络具有投影寻踪学习网络优点 ,在先验条件知道甚少的情况下 ,不用求点扩展函数 ,直接通过网络的学习 ,提取参数 ,以达到自适应剔除图像的模糊信息 ,恢复原图像 ;且具有小波函数的时域局部性 ,可以对多种噪声源的模糊图像进行恢复。模拟结果表明 。

Existing methods, including PPLN (projection pursuit learning network), are, in our opinion, still not quite satisfactory when used to restore blurred image without supervision. We combine PPLN with wavelet technique and propose PPWLN (projection pursuit wavelet learning network) for improving unsupervised restoration of image. Like PPLN, PPWLN trains the network by approximating degradation factors so as to bypass the difficult task of estimating point spread function when little is known about degradation factors. Unlike PPLN, PPWLN possesses the property of localization in the time domain; thus it can restore image, which is blurred by different noise sources, without supervision. The training time required by PPWLN is about half of that of PPLN. Figs.2, 3, 4, and 5 compare the unsupervised restorations of images blurred under various noise sources by PPWLN, PPLN, and traditional methods. We point out that Fig.5(c) restored by PPLN shows snowflake spots, while Fig.5(b) restored by PPWLN does not show spots. Table 1 gives the SNR (signal noise ratio) values for unsupervised restorations of blurred images under various noise sources by PPWLN, PPLN and traditional methods. These results indicate preliminarily that PPWLN is better than both PPLN and traditional methods in unsupervised restoration of blurred image.