Research On Image One-class Classification Algorithm Based On Deep Learning

Image classification aims to determine the category of the input image.It is one of the four basic tasks of computer vision.One-class classification is for the classification problem that contains only one category of samples in the data set.It is widely used in actual scenes.In recent years,with the development of deep learning theory,many excellent classification networks have been proposed to solve one-class classification problems.However,in practical applications,the classification effect of the classifier is often affected by many factors such as complex background,small data sample set,and small proportion of the target in the image.The result is not very satisfactory.The deep learning for one-class classification There is still room for improvement in performance.This article has conducted in-depth research around the above issues,the main research content and results include:(1)Constructed an ALOCC classification algorithm based on saliency detection.The saliency detection preprocessing operation is performed on the initial image data set.The processed image filters out the background or objects of the non-target subject,and only retains the target subject to be classified,which enhances the saliency of the target and improves the classification accuracy of the background features.The resulting interference phenomenon.The experimental results show that,without changing the network structure of the original ALOCC algorithm,using the saliency pre-processing proposed in this paper,and then sending it to the network for classification,the final classification result has been improved.In the public data set Caltech-101,Caltech-256 is verified,and compared with the experimental results of the original ALOCC algorithm,the AUC and F1 score values have been improved to a certain extent.(2)Constructed a one-class classification algorithm based on improved ALOCC network.Based on the original ALOCC network structure,combined with the AAE(Adversarial Auto Encoders)network,two adversarial generation networks are used for training,which improves the quality of the autoencoder's extraction of features in the training data set.When testing,call the autoencoder network that has been trained for the target category data set to complete the test on the data;on the basis of the loss function in the original network,the mean square error loss function of the encoding vector and the noise vector is added to enhance the network to the data the fitting effect.The single image classification algorithm constructed in this paper is verified on the public data sets MNIST,Fashion-MNIST and Caltech-256.Compared with the results of the original ALOCC algorithm,the classification effect has been significantly improved.Compared with the single classification algorithms in recent years,the overall performance has been improved to a certain extent.