Research On Graphical Reasoning Model In Visual Logic Based On MRNet

Visual logic of graphical reasoning reflects the ability to perceive connections between graphs.Graphical reasoning is the process of deducing a conclusion based on a number of known conditions consisting of a number of graphs as a premise with a certain law,with three elements: premise,conclusion,and reasoning requirements.Graphical reasoning does not depend on concrete things,and examines the ability of human observation,abstraction and reasoning,which is considered an important criterion to test human intelligence and thinking ability.Modern artificial intelligence can achieve performance beyond human level in some specific tasks,but the ability to reason logically through graphics is still far behind.One of the goals of artificial intelligence is to make machines possess human thinking,so making computers possess human-like reasoning ability is an important research component.Relational and Analogical Visual r Easo Ning dataset(RAVEN)provides structured datasets for graphical reasoning tasks and introduces graphical reasoning problems with different morphologies,and the logic of combinatorial transformations exists within the various morphologies.However,the current mainstream deep learning algorithm models are susceptible to the interference of attribute information that does not constitute inference rules when facing graphical inference problems with multiple combinatorial transformation logics and complex structures,which leads to the low inference ability of the models in graphical inference problems with some morphologies.In addition,the current mainstream model uses CNN for visual feature extraction,which cannot effectively retain shallow feature information,making the model lose more feature information before performing relational inference.In this paper,we address the problems of insufficient feature extraction ability and insufficient generalization,and aim to improve the inference ability of the machine by investigating the feature extraction,graphical inference model and other related techniques to achieve fast and accurate inference on RAVEN and Impartial-RAVEN(IRAVEN)graphical inference problems.The main contents of this paper are as follows:Firstly,this paper explains the background of tasks related to visual logic reasoning and the research significance of graphical reasoning tasks,introduces the current status of research on graphical reasoning problems at home and abroad,divides the existing methods for solving graphical reasoning problems into traditional methods and deep learning methods,and explains and analyzes the existing results respectively,and summarizes the advantages and shortcomings of the methods.Then we define and analyze the graphical inference task,briefly introduce the basics and theories of graphical inference,illustrate the deep learning neural network related techniques,the principle of attention mechanism and the variants of related attention mechanism,and introduce the representative datasets of graphical inference.Second,through analysis and comparison,this paper selects the Multi-scale Relational Network MRNet model,which has advantages in accuracy and time efficiency,and its end-to-end multi-scale relationship detection and independent scoring of individual options improve the robustness of the network model,but there are problems of insufficient feature extraction ability and poor inference under complex images.In this paper,we propose an improved graphical inference model based on MRNet,the Residual Attention Multi-scale Relational Network Res AMRNet model.For the lack of feature extraction capability,the backbone network is reconstructed using the residual structure,and a combination of hopping and long hopping is used to fuse shallow features into the deep network training to reduce feature information loss to improve the feature extraction capability of the model.For the poor inference ability of complex images,Dual-Pooling Efficient Channel Attention DECA is proposed to enhance the feature channel weights and reduce the interference,and fuse the channel attention mechanism with the residual module to detect the relationship features between each line of images in the inference stage,differentiate the importance of each feature channel In the inference stage,the channel attention mechanism is fused with the residual module to detect the relationship features between each row of pictures,differentiate the importance of each feature channel,learn the attention weight adaptively,extract the key features,and improve the graphical inference ability of the model under various morphological and combinatorial transformation.Finally,through experiments,we test the reasoning accuracy of the improved model,the result shows that the accuracy on RAVEN and I-RAVEN is 92.3% and 97.4%,respectively.Considering the drawback of generating candidate options on the RAVEN dataset,we used the I-RAVEN dataset standard dataset for the final ablation experiment,Under the same training sample and system environment setting,the Res AMRNet model proposed in this paper is compared with SRAN,Co PINet,DCNet and other advanced models to solve graphic reasoning.The results shows that Res AMRNet has obvious advantages in reasoning ability and reasoning speed on graphical reasoning tasks.Ablation experiments are also conducted to verify the improved module.The experimental results show that the addition of attention mechanism improves the reasoning ability of the model,the integration of DECA attention mechanism in the reasoning module can significantly improve the reasoning ability on complex graphs,and the improved backbone network also significantly improves the overall reasoning ability,which proves the effectiveness of the improved model.