西装识别的深度学习方法

Depth learning method for suit detection in images

为准确而快速地对电商平台产品图像进行西装目标的分类检测,以3个主要的卷积网络深度学习框架即快速区域卷积神经网络、基于区域的全连接卷积网络和单次多盒检测为基础,首先通过实验分析其在服装图像分类识别中的效率和有效性,针对小目标识别困难和过拟合识别问题,提出基于尺寸分割和负样本的单次多盒检测(SSD)增强方法(DN-SSD);然后将图像分割为不同尺寸的子图突出服装目标,通过融合分类方法解决SSD算法对小目标识别不足的问题,并通过增强负样本以提高算法的场景适应能力。实验结果表明,该算法可有效地识别各种形态和大小的西装目标,识别准确率达到90%以上,并且能够方便地推广到服装其他品类的识别中。

In order to classify and detect the suit target in images of e-commerce platform accurately and quickly,an enhanced deep convolution network (DN-SSD) was proposed.First,three main frameworks faster region-convolutional networks (faster R-CNN),region-based fully convolution network(R-FCN) and single shot muti-box detection(SSD) were evaluated.An image was segmented into multiscale sub-images to highlight the suit target based on the SSD.Secondly,the problem of small target recognition was solved by the fusion of classification.The scene adaptability was enhanced by increasing number of negative samples.The experimental result shows that the algorithm can recognize various shapes and size of suit targets and achieves the accuracy over 90%.The method can also be generalized to other style of dress detection and location.