Classification of scaled texture patterns with transfer learning

Abstract

Classification of texture patterns with large scale variations poses a great challenge for expert and intelligent systems. A pure learning approach addresses this issue by including texture patterns at all scales in the training dataset. This approach makes the construction of an expert system quite costly and unrealistic given the large variations in real-world texture scales and patterns. We propose a transfer learning approach where the full range of texture scales is available only for a small subset of the texture classes. Such a subset is used to learn the scaling map through partial least-square regression or coupled dictionary learning. Experimental results on classifiers equipped with the learned maps show promising reduction in training data scale variability with improved classification accuracy compared to the data-intensive pure learning approach. The proposed approach can be followed to build image-based expert systems of reasonable accuracy and limited data requirements.