Research And Improvment On Deep Neural Network Based Object Recognition

The artificial neural network can be well applied to image recognition task since its similarstructure with biological neural networks. At present, lots of exciting results were madeafter using deep structured neural networks into the image recognition task. However,there is still a long distance from using the deep neural network into practical applicationssuch as image retrieval, large-scale image annotation, graphic conversion, consideringthere are many problems to overcome. In this paper, the author did a detail study ofneural network structure, improved the training algorithm for deep neural network whichis commonly used, and proposed two new deep neural network model for differentapplication.Firstly, this paper did a detail study for the deep neural network based image recognitiontask, and proposed a series of algorithms to imporove the performance of the state-of-artmodel. The improved model obtained good results in the large scale visual recognitionchallenge. Based on this model, this paper designed and implemented an online imagerecognition system.During using the large scale users clickthrough data on the search engine to learnhigh-level image representation features, the author found the user clickthrough data isdifferent from the manually annotated training data, the clickthrough data is heavy taildistributed, and has large numbers of synonumous classes. These issues have a significantimpact on training multi-class neural network. This paper proposed a multi-task neuralnetwork which is insensitive to the data distributions. Singnificant gain on image retrievaltask was gotten after using the high-level image representation features learned from theproposed model instead of traditioanl low-level image representation features.At last, this paper proposed another deep neural network model base on Bag-of-wordsrepresentation called BoWDNN, which can learn visual based word embeddings and mapthe representation of images and words into same space. Finally, this paper built a systemsupports the similarity computation of word-to-image, image-to-image, word-to-word,and image-to-image.