Texture-Based Image Transformations for Improved Deep Learning Classification

In this paper, we examine the effect of texture-based image transformation on classification performance. A novel combination of mathematical morphology operations and contrast-limited adaptive histogram equalization is proposed to enhance image textural features. The suggested operations are applied in HSV colour space, where the intensity component is separated from the colour information. Two publicly available, texture-oriented datasets are used for evaluation in this study. The KTH-TIPS2-b dataset is utilised to illustrate the general effectiveness and applicability of the proposed solution on standardized texture images. The Virus Texture dataset is subsequently used to demonstrate a statistically significant classification improvement in a particular biomedical image recognition task.