| dc.description.abstract | The visually impaired people face many difficulties in daily life, especially in clothing identification and matching. In addition to having difficulty in independently knowing the color, pattern, style and other information of clothing fabrics, it is also a challenge for them to match each other well. Therefore, how to let the visually impaired understand the fabric information what they are wearing, and the suitability of the color and pattern of their clothes without the help of others, so as to avoid conflicts or unexpected clothes matching caused by insufficient information, is an important research topic. This research uses deep learning and image processing technology to develop an outfit improvement system in order to help the visually impaired understand the information about their clothes and improve the outfit collocation.
The deep-learning-based outfit improvement system developed in this paper mainly includes: (1) Skeleton detection: Use the OpenPose model to detect and remind the user whether the photo is a frontal full-body photo, (2) Human body part detection and segmentation processing: Obtain the human body mask through the CDCL Human-Part-Segmentation model, and perform focus cutting processing according to the corresponding parts of the clothing that need to be judged, (3) Clothing feature recognition: Use deep learning technology to train the model to recognize clothing information and fabric styles, (4) Recognition of fabric color system and saturation: Recognize the color information of fabric through image processing.
Through the functions of the above four parts, after obtaining clothing features and fabric color system and pattern style information, voice prompts will be used to assist the visually impaired to independently understand the features of each clothing they choose, and to give non-absolute wear reference suggestions to help them have a more conceptual and basis for the effect of clothing collocation and combination, while maintaining their flexibility in the beauty of collocation.
According to the system experiment results, the accracy of three kinds of clothing features and fabric styles has all reached more than 95%, the color classification accuracy rate is 87%, and the match between the recommended matching results and the taste of the general public is 77.6%, which can prove that the system has a certain degree of usability. | en_US |