Object detection of fashion apparel items is an indispensable technical condition for fashion industry and e-commerce. To address this need, this paper constructs an object detection system (YOLOv2-opt) based on the YOLOv2. In this work, we innovatively combine deep convolutional neural network with fashion apparel items detection. Compared with the traditional machine learning algorithms and semantic segmentation, the method based on deep neural network can locate and classify objects fast and better. Since the characteristics of fashion apparel detection, we propose an optimization for the input and network layers of YOLOv2 in order to improve the detection performance. The types of fashion items we consider in this work include trousers, skirts, coats, T-shirts, bags. Through the experiments, we evidence that our system has achieved an average precision of 0.839 and average recall rate of 0.73, detection speed reached 56ms per picture. Our optimized system is more exact than the YOLOv2.
Paper
Full text
An object Detection System Based on YOLOv2 in Fashion Apparel
Semantic Scholar · Computer Science · 2018
Abstract
Object detection of fashion apparel items is an indispensable technical condition for fashion industry and e-commerce. To address this need, this paper constructs an object detection system (YOLOv2-opt) based on the YOLOv2. In this work, we innovatively combine deep convolutional neural network with fashion apparel items detection. Compared with the traditional machine learning algorithms and semantic segmentation, the method based on deep neural network can locate and classify objects fast and better. Since the characteristics of fashion apparel detection, we propose an optimization for the input and network layers of YOLOv2 in order to improve the detection performance. The types of fashion items we consider in this work include trousers, skirts, coats, T-shirts, bags. Through the experiments, we evidence that our system has achieved an average precision of 0.839 and average recall rate of 0.73, detection speed reached 56ms per picture. Our optimized system is more exact than the YOLOv2.