Abstract: Image annotation based on deep features always requires complex model training and huge space-time cost. To overcome these shortcomings, an efficient and effective approach was proposed, whose visual feature was described by middle-level features of deep learning and semantic concept was represented by mean vector of positive samples. Firstly, the convolution result is directly outputted as the low-level visual feature by the middle layer of the pre-training deep learning model, and the sparse coding method was used to represent image. Then, visual feature vector was constructed for each textual word by the mean vector method of positive samples, and the visual feature vector database of the text vocabulary was constructed. Finally, the similarities of visual feature vectors between test image and all textual words were computed, and some words with largest similarities were selected as annotation words. The experimental results on several datasets demonstrate the effectiveness of the proposed method. In terms of F1-measure, the experimental results on IAPR TC-12 dataset show that the performance of the proposed method was improved by 32% and 60% respectively, compared to 2PKNN and JEC with end-to-end deep features.

Key words: deep learning, image annotation, convolution, mean vector of positive sample, feature vector