Deconvolution and Checkerboard artifacts,
A.Odena, A.Dumoulin and C.Olah, Oct.2016.
Neural Networks are used to generate images[GAN] starting from low resolution to give high resolution images which is done by deconvolution. While generating images deconvolution method creates checkerboard artifacts which is due to uneven overlapping positions. Sometimes kernels with even overlap can also develop this type of artifacts because deconvolution caneasily represent artifacts ceating functions even though the size is chosen carefully.
The artifacts can be avoided by choosing proper kernel size and stride. For even overlapping the kernel size should be divided by stride. The second approach is to resize the image by using Nearest Neighbor(NN)interpolation[approximates the value of a function] or bilinear interpolation[interpolation in 2 directions].By following the above 2 approaches the high frequency artifacts can be avoided.
Rather than using deconvolution, NN approach followed by convolution removes checkerboard artifacts. This paper also finds out that the artifacts are developed evn before training and it is because of the method employed. Computing gradients in the convolution layer, Max pooling can also cause high frequency artifacts which has to be attenuated.
The artifacts can be avoided by choosing proper kernel size and stride. For even overlapping the kernel size should be divided by stride. The second approach is to resize the image by using Nearest Neighbor(NN)interpolation[approximates the value of a function] or bilinear interpolation[interpolation in 2 directions].By following the above 2 approaches the high frequency artifacts can be avoided.
Rather than using deconvolution, NN approach followed by convolution removes checkerboard artifacts. This paper also finds out that the artifacts are developed evn before training and it is because of the method employed. Computing gradients in the convolution layer, Max pooling can also cause high frequency artifacts which has to be attenuated.