图像内容理解的深度学习方法

摘要

为生成准确描述视频内容的语句,研究计算机视觉领域图像识别的深度学习方法.基于广泛用于图像识别领域的端到端产生式模型,提出一种适合应用于视频内容理解领域的基于深层体系架构的产生式概率模型,建立将输入视频编码成向量,再将向量解码成完整句子的神经网络体系结构.通过在Sogou、mFlickr25k和MSCOCO图像数据集和网络视频数据集上训练的实验,分析生成的视频描述语句的语法准确性和语义准确性.实验结果表明,该产生式模型生成的句子比其它几个著名模型生成的句子获得了更高的BLEU和METETOR得分,验证了其有效性.%To generate sentences which can describe the video content accurately, the deep learning method of image recognition in computer vision realm was studied.Based on the end-to-end production model which was widely used in the field of image content understanding, a statistical model based on the deep schema which was suited for application in the field of video content description was presented.The neural network system structure for encoding the input video into vector and decoding the vector into sentences was put forward.The syntax and semantics accuracy of the generated video description sentences was analyzed through the experiment on Sogou, mFlickr25k, MSCOCO datasets and network video dataset.The result shows that the sentences generated using the proposed model achieve higher BLEU and METETOR scores than that generated by several other well-known models.The validity and efficiency of the presented method is demonstrated.