Unsupervised Feature Extration For Object Recognition

With the rapid development of information technology,image recognition is widely being applied to facial,digital and other object recognitions.So,strong technological needs exist in industrial,agricultural,commercial,military application.The key of image-recognition technology lies in feature extration for images on which the reasearches have important theoretical significances and application value.The paper carries out researches on usupervised feature extration method and applies it to the recognition of infrared vehicle targets.The main contents of the paper are as follows:This paper summarizes a variety of unsupervised feature extraction methods based on deep network,analyzes their advantages and disadvantages,discusses their limitations in the case of medium-sized data samples,and proposes the idea of unsupervised feature extraction that visual dictionary method takes the place of deep network.This paper proposes a single-stage unsupervised feature extraction method based on visual dictionary learning.After elaborating the process of unsupervised feature extraction,the influence of visual dictionary generation and encoder selection on feature extraction is discussed in depth.In this paper,we propose a combination of Kmeans clustering,OMP and soft threshold encoders to learn feature extraction from randomly sampled patches to improve the accuracy of recognition.On Cifar-10 and VOC2012 datasets,the improved single-stage computing structure is tested and compared with the feature extraction algorithm based on the deep learning.Experiments show that the recognition rate of improved single-stage computational structure is higher than that of the original single-stage computing structure and most other algorithms.It is verified that the proposed method is suitable for the application in case of medium scale training dataset.By simulating the convolution layer and pooling layer implementation method in deep neural network,and using spatial pyramid pooling method,we propose the single-stage expansion into two-stage computing structure to enhance its ability to extract more complex features.On self-selection dataset of remote-sensing buildings,the experimental results show that the two-stage computing structure is superior to the single-stage computing structure and other algorithms.This paper presents a method to detect the infrared identification of mobile vehicles from infrared images by combining the method of "objectness" and saliency detection with K-Feature unsupervised feature recognition.The results show that the method can effectively detect and recognize infrared mobile vehicles.The detection-recognition rate of single target sequence can reach 100%.In the case of particularly complex multi-target sequence,the detection recognition rate can be reaching about 78%.