Semi-supervised Image Classification Via Generative Adversarial And Self-attention

Computer vision is an important research interests in the field of artificial intelligence,which is to enable computers to observe the world like humans.In recent years,computer vision has been widely used in industries,transportation,medical care,military and other fields to improve people’s daily lives.Among them,image classification is one of the core algorithms in the field of computer vision,which is also the basis of other algorithms.The current method with higher accuracy in image classification methods is the supervised method based on deep learning.Although this method has high accuracy,it requires a large number of labeled samples.When the number of labeled images is too small,problems such as over-fitting may occur,thereby affecting the performance of the model.However,in practical application scenarios,images are often difficult to obtain or image labeling consume a lot of manpower and material resources,thus lack large-scale labeled images.Therefore,this paper mainly studies how to use a small number of labeled images and some unlabeled images to train a classification model with strong generalization ability.This paper constructs a semi-supervised image classification model based on generative adversarial network(GAN)and self-attention mechanism.First,this paper expands the discriminator from two-class to multi-class,and marks the generated images as fake samples to train the discriminator.Secondly,this paper introduces the self-attention mechanism in the improved generative adversarial network.The self-attention mechanism can calculate the correlation between each pixel and all pixels in the picture,and it can expand the receptive field.For generators,the self-attention mechanism can improve the quality of generated images,and high-quality generated images can improve the performance of semi-supervised image classification.For the discriminator,the self-attention mechanism can combine local features with global features to reduce the impact of the background.In addition,this paper also applies the spectral normalization technique to improve the stability of the model by performing a spectral normalization operation on the parameter matrix of the model.In this paper,manifold regularization is used as an additional regularization term to improve the robustness of semi-supervised classifiers to local perturbations.We tested the proposed method on the CIFAR-10 and SVHN datasets.The experimental results show that the method can achieve good results with a small number of labeled images.