数字式多媒体图像连续视觉特征标注仿真研究

Simulation Research on Continuous Visual Feature Labeling of Digital Multimedia Image

针对当前图像标注方法存在准确性低和实时性差的问题,提出基于深度特征与语义邻域的数字式多媒体图像连续视觉特征标注方法。根据MS算法将数字式多媒体图像划分成多个图像子块,计算各图像子块区域间权重值。利用NormalizedCuts分割算法融合各子块区域,获取图像分割结果。以图像分割为依据,训练图像语义分组,并利用逐层卷积与逐层采样法将图像抽象成特征向量,得到各语义组图像特征。将待标注图像也输入至已经训练好的深度网络,迭代提取特征步骤,计算待标注图像与各语义组中全部图像视觉特征相似程度,同时构建邻域图像集合。采用距离值法判断集合中各个语义标签贡献值,并对贡献值大小进行排列,得到预测关键词,即图像特征标注点,实现图像特征标注。实验结果表明,方法标注精度和效率均较高,具有可行性。

Due to low accuracy and poor real-time performance of current methods, this article put forward a method to label continuous visual feature of digital multimedia image based on depth feature and semantic neighborhood. According to MS algorithm, the digital multimedia image was divided into several image sub-blocks. The weights between sub-blocks were calculated. Normalized Cuts segmentation algorithm was used to fuse each sub-block region, so that the image segmentation results were obtained. Based on image segmentation, image semantic grouping was trained, and then the image was abstracted into feature vectors by layer-by-layer convolution and layer-by-layer sampling method, so that the image feature of each semantic group was obtained. The image to be labeled was also input into the trained depth network, and the feature steps were extracted iteratively. Moreover, the similarity among all visual features in the image to be labeled and each semantic group was computed. Meanwhile, the neighborhood image set was constructed. The method of distance value was used to judge the contribution value of each semantic label in set. In addition, the contribution values were arranged to get the predictive keyword, that is to say the image feature labeling point. Thus, the image feature labeling was achieved. Simulation results show that the proposed method has high labeling precision and efficiency.