Clothes Recognition and Annotation from Clothing Image Dataset

Abstract

Individuals view fashion as a significant form of self-expression that they use for a variety of purposes. It appears to be an essential component of every individual in contemporary society, covering both mundane activities and extraordinary experiences. Fashionable items are in high demand, making it a cherished and wealthy sector. In fashion, clothing items and their style play a prominent role in the selection of particular apparel. Conventional fabric identification techniques frequently depend on manual visual assessment, resulting in inefficiency and inaccuracy that fail to satisfy the requisite standards. Therefore, this creates a need and opportunity for the use and adoption of image recognition methods that utilize the latest computational intelligence techniques. Cloth image retrieval finds fashion images that match the query picture and meet user requirements. Due to the rapid rise of e-commerce and picture-sharing platforms and the enormous garment trade sector, academia and industry are focusing more on clothing research and application. Also, so many people buy things online, personalized picture searches are becoming more important, especially in the fashion industry. Because of the daily changes in this area, the clothing library grows quickly. Despite the high demand and research challenges, clothing-matching research and applications remain limited. Because of this, it is important to create a quick and useful method for identifying and retrieving pictures based on many things, such as color, pattern, material, and style. The existing feature extraction methods are effective at identifying cloth characteristics, but adding deep learning and other technologies makes the job easier and more useful. CNN-based methods are used a lot in cloth recognition these days. This thesis focuses primarily on the development and integration of DL-based methods for cloth image identification and retrieval. For e-commerce tasks like clothing recognition, clothing search, and clothing trait identification, the fashion industry uses a learning-based approach. But it's hard to recognize clothes because tags for clothes that look the same can be different. Therefore, this work proposed a framework is proposed for efficient apparel recognition by using a Convolutional Neural Network (CNN) and a Speed Robust Feature (SURF) together. A high-quality image of clothing gives more information, which leads to better feature extraction. So, the Atmospheric Dark Light Adjustment (ADLA) technique is proposed as a way to improve images during the framework's pre-processing stage. The effectiveness of the proposed integration (ADLA, SURF, and CNN) is validated by experimental results using six apparel attributes on three datasets, namely LSF, DeepFashion2, and Consumer to Shop. Further, the system is extended to retrieve similar images from the dataset for a given query image. The inherited quality of images is first enhanced through intensity modification and morphological operations achieved with the help of a light adjustment algorithm, followed by SURF feature extraction and CNN. Especially in the fashion sector, the proposed work is quite beneficial in generating accurate categorization and naturally appealing image retrieval from a library of images with diversified styles, patterns, and fashion. In addition to this, a modified approach named MFR-CNN, which is a modified Faster R-CNN approach, is proposed for faster cloth image retrieval. Using bounding boxes, the proposed model finds clothes and then sorts them into groups. Color, sleeve, collar, material, and dress type have been considered during the experimentation. The effectiveness of an algorithm is validated on cloth attributes by using an existing Deep Fashion dataset and also a new dataset created for the same manually. The proposed model detects clothes using bounding boxes and further classifies them. To determine the effectiveness, the proposed model is compared for accuracy, processing time, and total loss with three other methods, namely R-CNN, Fast R-CNN, and Faster R-CNN. The obtained results demonstrate that the proposed method has outstanding resilience as well as good efficiency in retrieving clothing images.