Accelerating The Detection Of Impulse Noise In Images With GPU

Image quality assessment is an important research in the field of digital image processing.The studies of blur-distorted image quality assessment algorithms and Gaussian noise-distorted image quality assessment algorithms have been appealed to most attention,while less researchers have focused on the methods of impulse noise-distorted image quality assessment.On the other hand,with the development of image sensor technology,digital images are several times in resolution and quantity compared with the past.It is very difficult for the traditional serial algorithms,which are designed to be executed on CPU,to meet the current real-time needs of massive images' processing.While GPU has become the preferred solution for parallel algorithms of high performance digital image processing due to its powerful computing performance.By analyzing the research status and development trend of image quality assessment algorithm,a GPU-based parallel quality assessment algorithm was designed for the impulse noise distorted images in this paper.Meanwhile,considering the application of massive image quality assessment system,an optimize strategy for the large-scale impulse noise-distorted images' quality assessment was also implemented.The main contents are described as following:(1)Based on the analysis of the impulse noise's characteristics and its filtering algorithms,a no-reference image quality assessment algorithm of the impulse noise distorted images was designed.The designed algorithm uses the gradient based structural similarity as the quality assessment index,which compared the difference of gradient based structural similarity between original corrupted image and impulse noise-free image which is constructed by a fuzzy filter algorithm.The experimental results show that the proposed algorithm can evaluate the quality assessment of impulse noise distorted images more accurately comparing with the full reference PSNR standard.(2)By analyzing the feasibility of the proposed serial algorithm,a GPU-based parallel quality assessment algorithm for the impulse noise distorted images was designed and implemented in this paper.In the parallel algorithm,CUDA kernels are designed considering the calculation process of serial algorithm and the characteristics of CUDA thread model and CUDA memory model.The experimental results show that the GPU-based parallel algorithm has much higher efficiency comparing with its serial counterpart under the basis of correctness in calculation result.Meanwhile,the GPU-based parallel algorithm achieves a maximum speedup of 65 X under the experimental environment with Tesla K20 c GPU.(3)Considering the application of massive image quality assessment system,an optimize strategy using CUDA stream mechanism was implemented to further improve performance of the GPU-based parallel algorithm.The optimize strategy regards each image's quality assessment process as one task,using the concurrency of CUDA stream mechanism to overlap data-transfer/data-computing and data-computing/data-computing between several tasks executing on GPU to exploit the capability of GPU device.The experimental results show that there exists at least 20% performance improvement can be obtained by the proposed strategy under the experimental environment with Tesla K20 c GPU.