基于深度学习的手绘草图识别

Sketch Recognition Using Deep Learning

现有的手绘草图识别方法严重依赖于费时费力的手工特征提取,而经典的深度学习模型主要是为彩色多纹理自然图像设计,用于识别手绘草图时效果不甚理想。提出一种基于深度学习的手绘草图识别方法(DeepSketch),该算法根据手绘草图缺失颜色、纹理信息的特点,使用大尺寸的首层卷积核取代自然图像识别中常使用的小尺寸首层卷积核,获得更多的空间结构信息。利用训练浅层模型获得的模型参数来初始化深层模型对应层的模型参数,以加快收敛,减少训练时长。加入不改变特征大小的卷积层来加深网络深度等方法以减小错误率。实验结果表明,所提出的方法较之其它几种主流的手绘草图识别方法具有良好的正确率,对250类手绘草图识别正确率达到69.2%。

In order to salve the existing problem of the sketch recognition heavily relying on the manual feature extraction which is very time-consuming, a method of sketch recognition based on deep leaming, called Deep-Sketch, was proposed. The classical deep learning models were mainly designed for natural color image recognition which failed on the sketch recognition. Deep-Sketch aimed to obtain more spatial structure information by using the large-size convolution kernel instead of the small-size convolution kernel in the first convolution layer. In addition, a shallow model was trained to obtain parameters which were used to initialize the corresponding layer parameters of the Deep-Sketch to reduce the model training time. Deep-Sketch was deepened with the convolution layers which kept the feature size to reduce the error rate. The results showed that the Deep-Sketch is superior to other state-of-the-art sketch recognition methods and achieves 69.2% accuracy on the sketch dataset including 250 classes.