CCD Noise Estimation And Reduction Based On Sparse Representation

Charge-coupled Device (CCD), the core component of a digital camera, will producecomplicated noise, which is not the simple additive Gaussian white noise (AWGN),but a kind of signal-dependent noise (SDN). Understanding CCD noise characteristicsare not only important for the quality assessment of image sensors, but also useful toadjust parameters of many computer vision algorithms. Therefore, it is quitesignificant for us to recover the noise level function (NLF) accurately.Most existed noise estimation algorithms assume that noise conforms to AWGNmodel, which cannot perform well for the signal-dependent noise. Therefore, wepropose a new method based on sparse representation in this paper, which needs toobtain three elements: a dictionary, noise estimation samples and their confidences.Firstly, we train a dictionary used for sparse representation by performing principalcomponent analysis (PCA) on a database, which is calculated in the consideration ofCCD types and their noise intensity of the real world. As for the noise samples andtheir confidence, they can be obtained by the noise estimation algorithm in the spatialdomain and the DCT domain, respectively. In the spatial domain, we use an imagestructure analyzer to detect homogeneous blocks and extract the CCD-output pixelsaccording to the pattern of color filter array (CFA) in order to calculate the noiseestimation samples. Meanwhile, the confidence of each sample is determined by thehomogeneity of the involved blocks. In the transform domain, we can only estimatethe noise samples and ignore the computation of their confidence because of theirhigh estimation accuracy. In this section, we perform the discrete cosinetransformation (DCT) on the down-sampled homogeneous blocks, and use thevariances of high-frequency coefficients to get the noise samples. After we obtain adictionary for the sparse representation, noise samples and their confidences, a CCDNLF can be recovered accurately based on the theory of sparse representation.After the noise level is estimated, we do further research on the CCD noisereduction. A new denoising algorithm is proposed in this paper, which is based onnoise estimation and3D-DCT. Experimental results show that our proposed methodaccurately estimates noise level functions for both smooth and highly-textured images. Moreover, our denoising method even can compare favorable with two state-of-the-artdenoising algorithms: Non-local Means (NLM) and BM3D.