The Fashion sector is among the most visually diverse and resourceful ones. Specifically, it requires the techniques of image processing and computer vision techniques in order to determine the widely prevalent fashion trends and public preferences, to generate advertisement policies, categorize apparel for online stores and search engines. It becomes imperative to generate models leveraging Computer Vision to classify this apparel and reconstruct lossy images that may arise from non-standardized sources while requiring little human intervention. This paper explores varied methods in Deep Learning that allow extraction of features in each image in the form of embeddings, pixel-wise segmentation of the apparel, followed by classification to the closest and most fitting class. The paper sets a base for a complete vision based solution for fashion items, which includes instance segmentation using Mask RCNN, feature extraction using embeddings and classification of all the fashion items present in a frame. Also, the performance of existing models are compared based on this dataset with thirteen main classes by ensuring minimum loss with the combination of certain hyperparameters.
Autoencoder Embedding, Segmentation and Classification of Apparel Images using a Deep ConvNet Approach
05.11.2020
397151 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Efficient ConvNet for real-time semantic segmentation
IEEE | 2017
|Efficient ConvNet for Real-Time Semantic Segmentation
British Library Conference Proceedings | 2017
|DepthCN: Vehicle detection using 3D-LIDAR and ConvNet
IEEE | 2017
|