Image denoising: Can plain neural networks compete with BM3D?
H.C.Burger, C.J. Schuler, S.Harmeling, IEEE conference on computer vision and pattern recognition, PP: 2392-2399
This paper tries to denoise an image by using Multi Layer Peceptron(MLP) and tries to compete with BM3D algorithm. A patch-based denoising algorithm is learned on a large training dataset with plain neural network. The clear input image is corrupted by adding Additive White Gaussian Noise(AWGN). MLP learns to map noisy image patches onto clean image patches. It is trained by using stochastic gradient descent and the parameters are updated using backpropagation algorithm. To make backpropagation algorithm more efficient, data normalization, weight initialization, learning rate division is done. Regularization is not applied to weights and no problems about overfitting.
The results are compared with denoising algorithms like GSM, KSVD and BM3D algorithms. MLP outperforms KSVD and GSM. MLP outperforms BM3D for 6 out of 11 images. BM3D algorithm is better for images with regular structure. Experiments were performed for several other noises like stripe noise, salt & pepper noise and JPEG quantization.
Software
The results are compared with denoising algorithms like GSM, KSVD and BM3D algorithms. MLP outperforms KSVD and GSM. MLP outperforms BM3D for 6 out of 11 images. BM3D algorithm is better for images with regular structure. Experiments were performed for several other noises like stripe noise, salt & pepper noise and JPEG quantization.
Software