Purpose. The purpose of image feature classification is to divide an image into several meaningful regions according to certain features, making these features the same or similar in a certain region but significantly different in different regions. In this paper, we will investigate the role of neural network and particle swarm optimization (PSO) in the image feature classification. Methodology. We propose the image feature classification method that combines PSO with neural network. BP neural network has been extensively applied in feature classification and it can classify specific objects or features through early learning, however, BP neural network algorithm also has many defects, including slow convergence speed and easiness to be trapped in local optimum. PSO optimized neural network fully exhibits its global search ability and parallel operation ability. Findings. Firstly, we take the gray image with specific object as the object to be segmented, study the samples with PSO neural network and get the training network. Secondly, we take the pixel matrix of the image as the input vector and put in the well-trained network for classification. Finally, the image feature classification can be realized. Originality. We made a study of image feature classification based on the particle swarm optimization neural network. We discussed the theory of image feature classification, basic principles of PSO and neural network. Practical value. We have also conducted the simulation experiment to confirm that the method suggested in this paper is a feasible one. We have proved that it has higher convergence speed and stronger robustness. Through the highly-efficient processing, this method can obtain important information and achieve excellent effect when used in the segmentation of the objects in complicated scenes.

Purpose. Financial statement fraud detection (FSFD) based on machine learning is a very important problem for avoiding financial risk and maintaining an orderly market. The purpose of this research was to develop a multiple instance learning model that is capable of detecting and predicting the risk of fraudulent financial reporting. Methodology. Each pair was composed of a singe-instance learning algorithm and its corresponding multiple instance learning algorithm, which were trained using a data set of 484 fraud companies as well as 902 normal companies with forming 4158 instances from Item 8 of the U.S. Securities and Exchange Commission (SEC) Form 10-K. Findings. Empirical study shows that MIBoost, miGraph and CKNN are superior compared to AdaBoostM1, SVM and KNN correspondingly in accuracy, F1 score and area under receiver operating characteristics curve (AUC), which prove that multiple instance learning algorithms can fit FSFD better, especially under class-imbalance and few training data. Originality. When a detecting label which corresponds to temporally local Financial Statement is attached collectively to groups of Financial Statements for one company without presenting the data to which Financial Statement this label is assigned, it is a multiple instance problem. The research presents a multiple instance learning model for FSFD originally. Practical value. We have also considered the fact that some auditors are dissatisfied with the single label learning algorithms because there are many instances in one company without label. Our model is more reasonable and accurate.