基于逻辑规则的图像内容理解和推理研究

抽象推理是人工智能的重要研究内容之一，其目标为从图像问题中进行特征提取和表示、隐藏规则分析及逻辑推理。尽管目前基于神经网络的计算机视觉已经在人脸识别、目标检测等任务中取得了巨大的进步，然而其抽象推理能力依然仍需要进一步提升。Raven’s Progressive Matrices (RPM)和人类的智力之间有着高度的相关性，是目前常用的智商测试工具之一。近年来，为了更好地了解和提升计算机视觉系统，RPM问题受到了越来越多的关注。本项目拟从多个不同角度展开基于神经网络的RPM抽象推理研究。（1）利用先验信息提取基于逻辑规则的视觉特征表示；（2）通过减小相同逻辑规则的视觉特征差异性，提升模型的泛化能力；（3）通过对多因素的深入分析，在保证推理模型准确度的同时提升模型的训练效率。本项目的研究将有力推动视觉图像理解和人工智能理论及应用的进一步发展。

Abstract reasoning is one of the most important topics in artificial intelligence, which focus on feature extraction and representation, hidden rules analysis and further logical reasoning on images. Although modern computer vision systems have made much progress on many tasks, such as face recognition and object detection, their reasoning ability is still needed to be further improved. Since Raven’s progressive matrices are highly correlated to human intelligence, it has been widely used for the IQ test of humans. Recently, to further understand and improve computer vision systems, solving RPM problems attracts more attention. This project will try to solve RPM problems from different perspectives, including follows. First, design logical rules based visual feature extraction with prior knowledge of RPM. Second, improve the generalization ability of models by reducing the diversity of visual features with the same logical rules. Finally, improve the accuracy and training efficiency by model optimization on multiple relevant factors. This research will greatly improve the theory and application of both computer vision understanding and artificial intelligence.