Saliency Detection And Denoising Algorithm For Noisy Images

Images are easy to be distorted by noise in the process of acquisition,transmission and storage.The existence of noisy images is very common.Noisy images not only affect visual experience,but also have a great impact on the digital image processing technology.Taking saliency detection as an example,this paper studies the saliency detection in noisy images,and proposes denoising method based on the characteristics of noisy images.Image saliency detection is useful in image compression,image segmentation,image retrieval and so on.Majority of existing saliency detection algorithms are presented for noise-free images.To solve this problem,this paper first evaluates the performances of state-of-the-art saliency detection algorithms using adapted image quality assessment database(Tampere Image Database,TID2013).TID2013 database has 24 distortion types which contains a variety of noise distortions.Since no saliency detection database exists for images with noises of different types and levels,we manually labeled the salient object in each image and get its ground truth image based on TID2013 database.The experimental results demonstrate that the distortions usually decrease the performance of the saliency detection algorithms,especially in high distortion levels.To solve the problem related to the decreased performances of saliency detection algorithms in noisy images,a machine learning-based framework for saliency detection in noisy images is proposed.First,we use the machine learning method to establish the mapping relationship between features of noisy images and noise levels,and obtain the noise level prediction model.Then we experiment and analysis the best denoising parameter setting of each noise type and level in noisy images.For each noisy image,the noise level prediction model is used to predict the noise level.And the noise is removed using the parameter setting corresponding to the noise type and level.Finally,the saliency map is calculated by using saliency detection algorithms and denoised image.Experimental results demonstrate that the proposed machine learning-based framework for saliency detection in noisy images improves the performances of saliency detection algorithms in most of the noise levels,particularly in high noise levels.In addition to image saliency detection,image noise affects the performances of image processing algorithms,such as image retrieval,image segmentation and so on.Therefore,the research of denoising algorithm is very important.The key of image denoising algorithm is to preserve the details of the original image while denoising the noisy image.A few of external denoising algorithms use clear natural images to retain more details of the image,but the use of external information needs the support of similar image regions.Most of image denoising algorithms are internal denoising algorithms which use the self-similarity of noisy images and use the same parameters to denoise the whole image.In this paper,we experiment with different denoising parameter settings and observe that different image regions prefer the denoising results obtained using different denoising parameter settings.In this paper,a machine learning-based region-aware image denoising algorithm is proposed.The final denoising result combines the denoising results obtained using different image denoising parameter settings in different image regions which can better retain the image texture details.The experimental results show that the proposed region-aware image denoising algorithm can effectively improve the performance of non local means(NLM),block-matching and 3D filtering(BM3D)and patch group prior based denoising(PGPD)algorithms and obtain higher Peak Signal to Noise Ratio(PSNR) and Structural Similarity (SSIM) values.