Blind people: clothing category classification and stain detection using transfer learning

Abstract

The ways in which people dress, as well as the styles that they prefer for different contexts and occasions, are part of their identity. Every day, blind people face limitations in identifying and inspecting their garments, and dressing can be a difficult and stressful task. Taking advantage of the great technological advancements, it becomes of the utmost importance to minimize, as much as possible, the limitations of a blind person when choosing garments. Hence, this work aimed at categorizing and detecting the presence of stains on garments, using artificial intelligence algorithms. In our approach, transfer learning was used for category classification, where a benchmark was performed between convolutional neural networks (CNNs), with the best model achieving an F1 score of 91%. Stain detection was performed through the fine tuning of a deep learning object detector, i.e., the mask R (region-based)-CNN. This approach is also analyzed and discussed, as it allowed us to achieve better results than those available in the literature.