Application Research Of Deep Learning In Image Classification

The accuracy of image classification affects the users' retrieval experience. The traditional image classification methods are suitable for processing a small amount of image data, and the classification accuracy is low. With the development of multimedia technology, the increasing image data has brought the image classification technique new challenges. The emergence of deep learning technology conforms to the requirements of big data era, which allows image features extraction not relying on the low-level features combination and transform which is greatly influenced by artificial factors any more. Deep learning technology can overcome the semantic gap between the low-level pixels and high-level semantic and it could imitate the human brain to extract image features. Deep learning model structure is complicated and various, including deep convolutional neural network, deep belief networks and so on. Deep convolutional neural network is widely used in image classification.Image can be divided into single-label image or multi-label image on the basis of image semantic complexity. Single-label image's semantic is simple so that an image only belongs to a category. While multi-label image's semantic content is rich, it belongs to different categories at the same time. To resolve single-label image and multi-label image classification problems, this paper deeply discusses the application of deep learning in image classification and the proposed method is introduced as follows:1. The improved depth image features extraction method based on low-level featuresDepth image features descript image high-level semantics, while the low-level features descript image color and pixels distribution regularity. Combined those two features, the paper proposes a new features extraction method called the improved depth image features extraction method based on low-level features. First, the method extracts image features using deep learning model. Then it extracts the color histogram features. Finally it generates a new features description method based on combination. The experiment uses Softmax and SVM classifier to classify the single-label image sets ImageNet and CIFAR-10, and the results show that the new features extraction method improves the precision of image classification. The method combines the low-level features and high-level features, which improves the ability of image features representation.2. The improved depth image features extraction method based on dimension reduction methodImage features dimension is still high after using deep learning technology to extract image features. This paper uses dimension reduction method to improve the depth image features extraction method so as to reduce the features dimension and complexity of the calculation. First, the method extracts image features using deep learning model. Then it uses two classic methods PCA and LDA dimension reduction methods to process depth image features, generating a new features description method. The experiment uses Softmax and SVM classifier to classify the single-label image sets ImageNet and the results show that the supervised dimension reduction method significantly improves the performance of image classification.3. The improved multi-label image classification bases on deep learningGenerally, image has multiple labels. Multi-instance multi-label learning is a method to solve the problems that exist in the multi-label image classification. The paper improves multi-instance multi-label KNN algorithm, including the bag generation method, the features extraction method and distance measurement method. Firstly, the method segments the image uniformly, and each image is evenly divided into four parts. Secondly, it extracts color-texture features and depth image features respectively, putting forward two methods for bag generating. Lastly, the method uses the minimum Hausdorff and average Hausdorff distance measurement for comparison based on multi-instance multi-label KNN algorithm. The experiment uses scene image sets and the results show that the depth image features are superior to the traditional color-texture features, significantly improving the scene image classification accuracy.