Research On Abnormal Image Detection Algorithm For The Financial Field Based On Visual Analysis

At present,the influence of Internet finance in China is continuously increasing.For Internet financial websites with a huge user scale,if they are attacked by hackers,and abnormal images such as pornography and violence are widely distributed,it will affect the normal operation of the website and cause certain economic losses.Therefore,it is necessary to detect and alarm at the source.Compared with manual filtering,the feature of deep learning to automatically extract features and classify makes the process of abnormal image detection more efficient.With the increase in model complexity,the decision-making process inside the model becomes difficult to understand,and the direction of model improvement is gradually blind,which makes it more and more important to establish an intuitive understanding of the model building process.Aiming at the characteristics of Internet financial websites that have high requirements for information sensitivity,this thesis establishes an abnormal image detection model for the financial field based on visual analysis technology.Specific work is as follows:(1)An analysis algorithm for abnormal image detection model is proposed.First,normal images,pornographic images and violent images for the financial field are collected and preprocessed to form a dataset for model training.Pornographic image detection model and violent image detection model are built applied the Res Net-50 network based on this dataset,and the experiment and result analysis are completed.The experimental results show that the performance of the two models is relatively close;Secondly,according to the characteristics of the Res Net-50 network structure,the two detection models are used as the analysis object,and the image channel is used as the research granularity.The channel importance algorithm in the model layer and the channel influence algorithm between the model layers are proposed to understand the decision-making process of the model.The matrix output by the two algorithms can be combined to form a graph structure;Finally,Personalized Page Rank Algorithm is used to simplify the graph structure,and the key structure is extracted to form a model attribution graph.(2)Visual analysis of abnormal image detection model is carried out.First,the model attribution graph obtained from the analysis of the abnormal image detection model is used as the data source for visual analysis,and the design of the visual view is completed from the perspectives of overall structure,channel importance and channel influence;Secondly,the model is understood as the target,the network is the center,and the node-link diagram is the layout,and the view construction is completed.Finally,based on the constructed view,considering the importance of the channel and the influence of the channel,the selection criteria for the key channels in the model are formulated through visual analysis.(3)An abnormal image detection model based on visual analysis is constructed.The SE block in the channel attention mechanism is used as the basis,and the channel importance information obtained by the Excitation operation and the channel importance information obtained by visual analysis are combined in this structure.In order to analyze the effect of the detection model,this thesis carried out experiments on the pornographic dataset and the violence dataset,and compared the detection performance of the model based on Res Net-50,the model based on SE-Res Net-50 and the model based on visual analysis.The experimental results show that the accuracy,precision,recall and F1-score of the model based on visual analysis have been improved on the two datasets,which reflects the use of visual analysis technology to intuitively understand the model learning process,which is conducive to the construction of the abnormal image detection model for the financial field to better meet the needs of abnormal image detection for the financial field.