Hand-designed local image descriptors vs. Off-the-shelf CNN-based features for texture classification: An experimental comparison

Convolutional Neural Networks have proved extremely successful in object classification applications; however, their suitability for texture analysis largely remains to be established. We investigate the use of pre-trained CNNs as texture descriptors by tapping the output of the last fully connected layer, an approach that has proved its effectiveness in other domains. Comparison with classical descriptors based on signal processing or statistics over a range of standard databases suggests that CNNs may be more effective where the intra-class variability is large. Conversely, classical approaches may be preferable where classes are well defined and homogeneous.