基于深度学习的服装三要素识别

Recognition of Clothing"Three Elements"Based on Deep Learning

为快速自动获取服装三要素信息,提高服装图像多特征识别效率,提出一种利用深度学习识别服装三要素的方法。考虑款式、颜色、图案3种要素,建立了一个包含3种上衣款式、6种颜色、6种图案,共计15种类别的样本库,利用改进的VGGNet神经网络进行款式与颜色识别,结合YOLOv3,Faster R-CNN,SSD目标检测算法实现图案识别及定位。对比实验结果,得出改进的VGGNet对服装款式与颜色识别准确率达到96.49%;目标检测算法中YOLOv3对服装图案识别与定位的mAP达到86.66%,3大类图案中纹理类图案的检测效果最好,其mAP为96.14%,动物类图案mAP为83.69%,文字类图案mAP为79.80%。研究结论为顾客服装偏好信息的快速获取提出了新思路。

In order to quickly and automatically obtain the information of the three elements of clothing and improve the efficiency of multi-feature recognition of clothing images,a method for identifying the three elements of clothing using deep learning was proposed.Considering the three elements of style,color and pattern,a sample library was established including 3 tops styles,6 colors and 6 patterns,a total of 15 categories.It used the improved VGGNet to identify colors and styles,and combined with YOLOv3,Faster R-CNN and SSD target detection algorithms to achieve rapid pattern recognition and positioning.The comparative experimental results show that the improved VGGNet has an accuracy of 96.49%for clothing style and color recognition,and the YOLOv3 in the target detection algorithm has a mAP of 86.66%for clothing pattern recognition and positioning.Among the three types of patterns,texture patterns have the best detection effect.Texture patterns'mAP,animal patterns'mAP and text pattern's mAP are 96.14%,83.69%and 79.80%respectively.This study puts forward a new idea for the rapid acquisition of customer clothing preference information.