Texture Recognition Using a Novel Input Layer for Deep Convolutional Neural Network

We introduce a new type of layer as input to the convolutional neural network. The layer is designed to capture pixel to pixel relational properties that are found in textures. These pixel relational properties are not easily captured in a conventional convolutional neural network (CNN) with RGB (red-green-blue) layer input. This approach uses the CNN architecture with a pre-input processing to capture the pixel relational properties such as the correlation between the center and surrounding pixels. We tested our approach against multiple texture datasets such as Colored Brodatz, Multiband Texture and Wood texture. The results are competitive against current best results, indicating that this approach is promising approach.