Deep Learning-based Fast Denoising Algorithm For Random Impulse Noise

So far,the performance of most denoising algorithms for random valued impulse noise(RVIN)are seriously limited by whether to accurately detect the noisy pixels in the image to be denoised.Therefore,it is great theoretical significance and practical value to propose an algorithm that can quickly and accurately detect the noisy pixels in a noisy image.As the preprocessing module of most denoising algorithms,the detection accuracy and execution efficiency of the denoising algorithm are two important indicators reflecting its performance.Most existing RVIN algorithms adopt method that detection and denoising execute serial iteratively pixel by pixel,which lead to the execution is inefficient.In addition,the accuracy of detection algorithms that comparing the hand-crafted local image statistic(LIS)with the preset threshold to detect the noisy pixels need to be improved.To solve the problems of denoising algorithms on detection accuracy and execution efficiency,two fast RVIN detection algorithms were proposed:1)A RVIN detection algorithm based on deep belief network(DBNRD).It was implemented by constructing a more descriptive feature vector and training a detection model with more accurate nonlinear mapping.On the one hand,;multiple rank-ordered logarithmic absolute deviation(ROLD)statistics were extracted and combined with a statistic about the edge feature in the form of feature vector to describe how RVIN-like the center pixel of a patch is.The description ability of the feature vector is improved significantly while the computational complexity is just increased in small amount.On the other hand,an RVIN prediction model(RVIN detector)was trained by DBN to map the feature vectors to noise labels,which is more accurate.Once the RVIN detector was trained,the noise detection can be completed in one step,which achieves significant practical value.2)A RVIN detection algorithm based on convolutional neural network(CNNRD).Although DBNRD algorithm can achieve the expected results,it is still a traditional algorithm based on the hand-crafted extraction of feature for detection.Therefore,we adopted the convolutional neural network(CNN)with powerful feature extraction function to train the RVIN detector based on image patches.The detection model directly input raw patches rather than artificial features such as ROAD and ROLD,so that CNN can automatically extract the features indicating noise pixels.In addition,to simplify the network structure and improve the execution efficiency,we disposed training patches properly.First,the difference value between and its neighboring pixels were sorted by ascend.Then the sorted data was recombined into a new patch as the final input of the network.CNNRD algorithm avoids the limitation and complexity of manually designed LIS statistics and thresholds,and noise detection and denoising need not execute serial iteratively.In addition,the detection model based on CNN network greatly improves the efficiency of CNNRD,especially under the condition of GPU.To cooperate with DBNRD and CNNRD for quickly and effectively completing the following denosing,We proposed a deep CNN-based RVIN denoising algorithm(DCNNRDA).The existing switched RVIN denoising algorithm according to the detection results restore the noisy pixels pixel by pixel,which has inefficient execution.The end-to-end denoising model,trained directly on a large number of data with deep learning,which is based on deep learning and trained directly on a large number of data,shows a higher efficiency than the switched algorithms.However,this kind of denoising algorithm depends on data seriously.When the RVIN level of noisy image is different from the image set used in the training denoising model,the performance of model will be reduced to some extent.To solve the problem of data dependence in data-driven denoising algorithm,we divided the number of the detected noisy pixel by the total number of detected pixels to convert it into noise ratio.We proposed a fast non-switched RVIN denoising algorithm which consists of two modules,i.e.,RVIN detector and denoising.Specifically,in the training stage:Firstly,the noise ratio range(0%-60%)was divided into several sub ranges.Then,denoising models were trained in each sub range.In the test phase:For any given noisy image,DBNRD or CNNRD is applied to detect noisy pixels,then the number of detected noisy pixels divide by the total number of image pixels converting into noise ratio to measure the level of RVIN.Finally,according to the noise ratio,denoising model is used adaptively to remove the RVIN from the noisy image.The denoising algorithm belongs to the non-switched,which makes its execution efficiency significantly higher than the classical RVIN denoising algorithms,and the deep CNN network structure ensures the denoising effect.To assess the performance of DBNRD,CNNRD and DCNNRDA,we compared them with state-of-the-art denoising algorithms in terms of detection accuracy,execution efficiency and denoising effect on common used image set,BSD and Waterloo database.A large number of experimental results show that the proposed noise detectors achieve better comprehensive performance on detection accuracy and execution efficiency compared with state-of-the-art RVIN detectors.In addition,the denoising effect of the corresponding pre-trained DCNNRAD model achieves significant advantage.