Faster Image Annotation Based On Multi-Label Deep Learning For Weakly Labeled Dataset

With the rapid widespread of digital techniques, the number of images has been rised rapidly. How to manage these images and get the useful information from these images is an urgent problem. The multi-label annotation of image is a good solution by using natural language to describe the image so that it has become one of the hottest research points. Now there are many problems with the task of multi-label image annotation:The feature extraction of traditional machine learning algorithm is very complex and could not achieve the end-to-end output, which badly affects commercial applications. Although the deep learning has better result of multi-label image annotation, but this method is too slow for commercial application because of the deep learning model has a large number of parameters.What is more, the training data sets are always not completely, and has the problem of semantic distribution imbalance in real world, which named weakly labeled data sets, would badly affects the result of image annotation. In view of the above problems, this paper proposes a method of faster image annotation based on multi-label deep learning for weakly labeled dataset. The main work of this thesis is as follows:Put forward a method of faster image annotation based on multi-label deep learning. This thesis put a new model which has fewer parameters by modifying the stride of the Convolution kernel and the number of outputs of each layer. what is more, this thesis has been decomposed some convolution layers by SVD so that the new model runs almost 6 times faster than VGG-16 model and the macro-accuracy of model just reduced 2.5%.Put forward a iterative optimization algorithm framework for the weakly labeled dataset. The iterative optimization algorithm framework through adding the single labeled images of the low-frequency semantic label to improve the recognition accuracy of the low-frequency semantic. what is more, the mothed of feature fusion and KNN classifier was used in the prediction step. Finally, The number of tags increased 7% in Tencent dataset and the macro-accuracy of the new data set increased 1.2%.These work has been applied in Image Understanding of Tencent BestImage Open platform and WeiYun of Tencent, achieve good commercial application.