Concerning the problem of semantic gap between low-level visual features and high-level semantics in automatic image annotation, based on traditional dictionary learning, a multi-label discriminative dictionary learning method was proposed to automatic image annotation. First of all, multiple types of features for each image were extracted, and a combination of a variety of features was used as input information of the input feature space to the dictionary learning. Then, a label consistency regularization term was designed to integrate the label information of the original samples into the initial input feature data, and the dictionary of label consistency and the label consistency regularization term were combined to learn the dictionary. Finally, the label sparse coding vector was obtained by the dictionary and sparse coding matrix to implement the semantic annotation for an unknown image. The performance of the annotation was tested on the Corel 5K data set. The average precision and average recall could reach 35% and 48% respectively, compared with the traditional Sparse Coding Method (MSC), which were increased by 10 percentage points and 16 percentage points respectively, and increased by 3 percentage points and 14 percentage points respectively than the method of Distance Constraint Sparse/Group Sparse Coding (DCSC/DCGSC) for automatic image lableing. Compared with the current image annotation methods, the experimental results show the proposed method can predict the semantic information for an unknown image properly, and has better annotation performance.

The sparse network structure for traditional Convolutional Neural Network (CNN) can not preserve the high efficiency of dense network-intensive computing and the empirical selection of the activation function in the experiment process, which leads to inaccurate results or high computational complexity. To solve above problems, an improved CNN method was proposed and applied in remote sensing images classification. Firstly, the multi-scale features of an image was extracted by using different scale convolution kernels of the Inception module, then the activation function of the hidden layer node was studied by using the Maxout model. Finally, the image was classified by the Softmax method. Experiments were conducted on the same US Land Use Classification Data Set 21(UCM_LandUse_21), and the experimental results showed that the accuracy of the proposed method was about 3.66% and 2.11% higher than that of the traditional CNN method and a Multi-Scale Deep CNN (MS_DCNN) respectively with the same number of convolution layers, and it was also more than 10% higher than that of visual dictionary methods based on low-level features and middle-level features. The proposed method has high classification efficiency and is suitable for image classification.