基于深度学习的酒标分割研究

红酒图像中的酒标区域含有重要的红酒品类信息,而对酒标区域的定位与分割可以有效去除背景区域对图像匹配算法的干扰。传统图像分割算法大多基于局部图像特征和人工设计规则,对噪声较为敏感,并且难以应对大规模数据的处理。针对传统算法的不足,本文首先构造了一个大规模酒标分割数据集,然后提出了一种基于深度学习的酒标分割方法。我们构造了一个基于残差网络的语义分割模型,并且在模型中加入跨层连接,实现低层特征和高层特征的有效融合,使得分割的边缘细节更加清晰和准确。另外,我们采用了带孔卷积金字塔池化结构整合多尺度信息,在增大模型感受野的同时使得模型适应不同尺度的酒标区域。在我们构造的酒标数据集上的实验结果表明,本文提出的酒标分割网络能够进行实时的酒标图像分割,并且达到了相当高的分割准确率。

Research on Deep Learning Based Wine Label Segmentation

The label area of a wine bottle contains its identification, while wine label segmentation can effectively eliminate the interference of the background in the image matching algorithm. Most of the conventional image segmentation algorithms are based on low level features and human defined rules, which are sensitive to the noise and have trouble in processing massive data. In this paper, we first construct a large scale wine label segmentation dataset, then propose a wine label segmentation method based on deep learning to tackle the problems in the conventional algorithm. We design our semantic segmentation model based on a deep residual network with certain skip-layer connections which integrate the low-level and high level features together to achieve clearer marginal details of segmentation. Furthermore, we adopt Atrous Spatial Pyramid Pooling(ASPP) to enlarge the receptive field while segmenting multi-scale wine labels. It is demonstrated by the experimental results on our wine label dataset that the proposed wine label segmentation algorithm can achieve high accuracy in real-time.