目的 基于卷积神经网络(CNN)在图块级上实现的随机脉冲噪声(RVIN)降噪算法在执行效率方面较经典的逐像素点开关型降噪算法有显著优势，但其降噪效果受制于对待降噪图像受噪声干扰程度(噪声比例值)的准确估计。为此，本文提出一种基于多层感知网络的两阶段噪声比例预测算法以自适应地调用CNN预训练降噪模型达到获得最佳去噪效果的目的。方法 首先，对大量无噪声图像添加不同噪声比例的RVIN噪声构成噪声图像集合；其次，基于视觉码本(visual codebook)采用软分配(soft-assignment)编码法提取并筛选若干能反映噪声图像受随机脉冲噪声干扰严重程度的特征值构成特征矢量；再次，将从噪声图像上提取的特征矢量及其对应的噪声比例分别作为多层感知网络的输入和输出训练噪声比例值预测模型，实现从特征矢量到噪声比例值的映射(预测)；最后，采用粗精相结合两阶段实现策略进一步提高RVIN噪声比例的预测准确性。结果 针对不同RVIN噪声比例的失真图像，本文从预测准确性、实际降噪效果和执行效率三个方面来验证所提出算法的性能和实用性。实验数据表明，本文算法在大多数噪声比例下的预测误差小于2%，应用于CNN降噪算法的降噪效果(PSNR指标)较其他主流降噪算法高2~4dB，处理一张 大小的图像仅需3秒左右。结论 本文提出的RVIN噪声比例预测算法在各个噪声比例下具有鲁棒的预测准确性，依据其所预测的比例值调用预先训练的CNN降噪模型所实现的RVIN降噪算法在降噪效果和执行效率两个方面较经典的开关型RVIN降噪算法有显著提升，更具实用价值。
Two-Stage Multi-Layer Perceptron Estimation For Random-Valued Impulse Noise Ratio
Yu Haiwen,Yi Xinwei,Xu Shaoping,Zhang Guizhen,Liu Tingyun(School of Information Engineering,Nanchang University,Nanchang,330031)
Objective The existing switching random-valued impulse noise (RVIN) removal algorithms mainly detect the noisy pixels of an image to be denoised by comparing the local image statistic with predefined thresholds, and then combine a denoising method to restore the detected noisy pixels in a pixel-wise manner, resulting into low execution efficiency. With respect to computational complexity, the convolutional neural network (CNN)-based denoising algorithms that were implemented at patch-level for random-valued impulse noise (RVIN) has a significant advantage over the classical switching denoising algorithms that detect and remove RVIN pixel-by-pixel. However, the restoration performance of the CNN-based denoising algorithms is still limited to the accurate estimation of the distortion level of the given noisy image. In essence, the CNN-based denoising algorithm is still a non-blind method, where the best denoising effect only can be obtained by training a specific denoising model at a fixed noise level, limiting the practical application. For simplicity, the noise ratio can be treated as a measure of the distortion level of a noisy image, by dividing the number of detected noisy pixels by the total number of image pixels. According to the estimated noise ratio, CNN-based denoising methods can remove the RVIN blindly and efficiently with high quality by exploiting the corresponding pre-trained denoisers adaptively. To precisely estimate the noise ratio, a two-stage noise ratio estimation algorithm based on multi-layer perceptron (MLP) was proposed in the paper. Method Specifically, a large number of clean images were first corrupted with RVIN at different ratios to form a set of noisy images. Then, based on the visual codebook and soft-assignment coding technology, the features that can reflect the distortion level of a noisy image were extracted and screened to form feature vector for each noisy image. After that, the feature vectors and their corresponding noise ratios extracted from noisy images were used as the input and output of the multi-layer perceptron model respectively to train the noise ratio estimation model that maps a given feature vector to its corresponding noise ratio. Generally, to obtain the ideal approximation function, the more hidden layers are required in MLP architecture. But the construction of MLP-based regression model with multi-hidden layers is difficult in convergence and the training speed. Therefore, a coarse-to-fine two-stage strategy was employed to further improve the estimation accuracy. Concretely, a relatively coarse noise ratio estimation model was trained across the whole range of noise ratio, and then the noise ratio range was divided into a number of sub-ranges, which means the mapping range of the estimation model is reduced. In the same way, several fine noise ratio estimation models were trained in different noise ratio sub-ranges. Note that each subinterval overlaps with its adjacent subinterval to avoid the estimation inaccuracy at the extremities of subinterval. In the prediction phase, a preliminary estimation is first obtained using the coarse estimation model. Based on this, the corresponding fine estimation model is employed to predict the noise ratio more accurately. Result The comparison experiments were conducted to test the validity of proposed method from three aspects: estimation accuracy, denoising effect, and execution efficiency. The proposed method was first compared with several classical noise detectors of RVIN denoising methods, such as PSMF, ROLD-EPR, ASVM, and ROR-NLM, to demonstrate the estimation accuracy. The number of detected noisy pixels was converted into noise ratio, since the output result of the noise detectors of those compared switching denoising methods is the number of noisy pixels. Results shows that the estimation error of the proposed method is less than 2% across different noise ratios, showing stronger robustness than others. To verify the availability of the proposed method, the feed-forward denoising convolutional neural network (DnCNN) algorithm that is designed for removing Gaussian noise was improved to deal with the removal of RVIN. In denoising effect comparison, the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and feature similarity (FSIM) were adopted as image quality assessment index. For the distorted images with different RVIN noise ratios, the PSNR values obtained by the improved DnCNN algorithm utilizing the proposed method increase by 2 dB more than that of others across the range of noise ratio from 10% to 60%. Simultaneously, the FSIM values rank in the top 2 for different noise ratios, while the SSIM values approximate the best results. Regarding qualitative visual evaluation, the improved DnCNN algorithm utilizing the proposed estimation model can generate a clearer restored image with better edge preservation. Compared with switching RVIN removal methods, the improved DnCNN algorithm outperforms them in execution efficiency, which takes only 3.8 seconds to restore an image of size . Conclusion Extensive experiments show that, the estimation accuracy of the proposed MLP-based noise ratio estimation algorithm is robust across a wide range of noise ratios. With the proposed noise estimation model, the CNN-based RVIN removal algorithms can achieve the best blind denoising by exploiting the closest matching model. Moreover, compared to the traditional switching RVIN denoising algorithm, the improved DnCNN denoising algorithm with the noise ratio estimation module outperforms them significantly in terms of both denoising effect and execution efficiency, which makes it more practical.