基于目标检测算法的FashionAI服装属性识别

随着网络上服装图片数量的快速增长，对于大量的服装进行分类的需求与日俱增.传统的使用手工进行服装图像的语义属性标注并不能完全的表达服装图像中的丰富信息，并且传统的手工设计的特征已经不能满足现实的精度和速度的需求.近年来，深度学习已经应用到计算机视觉方方面面，为基于深度学习的服装分类识别技术奠定了坚实的基础.本文根据已有的数据集DeepFashion构建了三个新的子数据集，进行分类训练的deepfashionkid数据集和进行Faster R-CNN训练的deepfashionVoc数据集和进行Mask R-CNN训练的deepfashionMask数据集.使用deepfashionkid数据集在VGG16上进行预训练得到clothNet模型，进而改进Faster R-CNN的损失函数.并且各自对比了这两种算法使用clothNet预训练的模型与不使用的区别.另外，本文了采用一种新的类似嫁接学习的预训练策略.实验表明，这些训练技巧对于检测精度的提高具有一定的帮助.

FashionAI Clothes Recognition Based on Object Detection Algorithm

With the rapid growth in the number of clothing pictures on the Internet, the demand for classification of a large number of clothing is increasing. The traditional use of manual semantic attribute annotation of clothing images does not fully express the rich information in the clothing image, and the traditional hand-designed features can no longer meet the requirements of real precision and speed. In recent years, deep learning has been applied to all aspects of computer vision, laying a solid foundation for clothing classification and recognition technology based on deep learning. In this study, three new sub-datasets are constructed according to the existing dataset deepfashion, the deepfashionkid dataset for classification training, the deepfashionVoc dataset for training with Faster R-CNN, and the deepfashionMask dataset for Mask R-CNN training. The clothNet model is pre-trained on the VGG16 using the deepfashionkid dataset to obtain the clothNet model, which in turn improves the loss function of the Faster R-CNN. And each compares the difference between the two algorithms using clothNet pre-trained model and not used. In addition, this study adopts a new pre-training strategy to adopt a training method similar to grafting learning. Experiments show that these training techniques are helpful for improving the detection accuracy.