Research On Self-supervised Learning Method Based On Pre-training Feature Extraction And Its Applications In Image Recognition

At present,in the era of rapid development of big data,with the great improvement of computing resources such as cloud computing and high-performance computing,deep learning model can be used to store massive data and make more accurate prediction and analysis of future or unknown events through strong learning ability and efficient feature expression ability.In this context,deep learning has emerged and made good progress in speech recognition,image recognition and natural language processing.Among them,with the emergence of large-scale annotated data sets,deep learning has achieved critical success in the field of computer vision.However,collecting data sets and manually labeling of data requires significant labor costs,and a lot of unlabeled data is now easily accessible from the network.Self-supervised learning can learn distinguishable visual features by designing auxiliary tasks,identify labels from a large number of training data or images,and provide supervision basis for computer vision research model training.In the era of artificial intelligence,computer vision mainly focuses on image recognition.Image recognition mainly includes two steps: feature extraction and classification recognition.Different from traditional manual feature extraction,deep learning algorithms can automatically extract features.Some images contain single information,while others contain multiple objects with complex background.This paper proposed two image recognition algorithms for single-label image recognition and multi-label image recognition:In terms of single-label image recognition,for the processing of unlabeled or a small amount of label data,this paper proposed an image classification algorithm named OCFC based on self-supervised learning without manual labeling.The main idea of the algorithm is as follows: first,after preprocessing the image with the OTSU segmentation algorithm,three-layer Convolution Restricted Boltzmann Machine extracts features.Then the extracted feature clusters are labeled with pseudo-labels by the Fuzzy C-means algorithm.Finally,use the CNN model for classification and recognition and predict the category of other images.The self-supervised learning model can be arbitrarily transferred to a shallow model or a deep model.The experimental results show that this method can effectively avoid the complexity of manually labeling data,and the accuracy on the STL-10 data set reaches 82.7%,and theaccuracy on the CIFAR-10 data set reaches 91.2%.In terms of multi-label image recognition,this paper proposes a multi-label image recognition algorithm of DCGAN and GCN based on Self-supervised learning.We establish a dependency model between each object label in the image through a graph neural network.At the same time,we use a self-supervised learning method based on the generated data pre-training model method to improve the recognition accuracy of multi-label images.In this paper,Deep Convolution is used to generate DCGAN model to generate similar images to increase the training data of the data set,so as to alleviate the robustness of the over-fitting model in the training process.Then the directed graph between the object labels in the image is established and map the class labels to the corresponding class classifiers by the Graph Convolutional Network(GCN).Then the parameters of the Convolution Neural Network of the DCGAN model that generated the data were transferred to the CNN model for the feature extraction of multiple label images.Next,the extracted feature vector is multiplied with the matrix of the classifier generated by GCN training to obtain the vector for classification,and finally,the multi-label images are classified and recognized.The algorithm has carried out related experiments on two multi-label data sets,and the experimental results show that the algorithm is superior to other multi-label image recognition methods in the two data sets.Based on the Self-supervised learning method,this paper studies the problem of single-label image recognition and multi-label image recognition.It proposes corresponding algorithm models,which improves the performance of image recognition algorithms in corresponding fields and has specific theoretical and application value.