A Non-Linear Convolution Network for Image Processing

This paper proposes a new neural network structure for image processing whose convolutional layers, instead of using kernels with fixed coefficients, use space-variant coefficients. The adoption this strategy allows the system to adapt its behavior according spatial characteristics input data. type layers performs, as we demonstrate, non-linear transfer function. features generated by these compared ones canonical CNN are more complex and suitable fit local images. Networks composed offer performance comparable or superior which Convolutional Networks, fewer significantly lower number features. Several applications newly conceived networks classical image-processing problems analyzed. In particular, consider: Single-Image Super-Resolution (SISR), Edge-Preserving Smoothing (EPS), Noise Removal (NR), JPEG artifacts removal (JAR).