Application Of Deep Learning In Image Semantic Classification

Image classification has been the hot-spot problem in academic research, and also the foundation of image indexing, image understanding and analysis, et al. In recent years, with the increasing of digital images in the Internet, traditional image classification methods failed to satisfy the changing demands of practical applications, which thus drives people’s growing concerns for the image classification methods based on semantics. The essay dwells on proposal for major characteristics of which hierarchical semantic features must satisfy through further analysis of the latest research results regarding biological vision and image features. Moreover, the proposal for applying deep learning models to learn hierarchical semantic features of images, which is on the basis of the ability of deep learning models to learn hierarchical features from data. Furthermore, we designed two deep learning models to implement semantic classification of image.The essay dwells on such aspects as follows:(1) Introduction on the significance of image semantic classification and general thoughts to deal with image semantic classification, discussion on the two types of methods for constructing image semantic features, evaluation about disadvantages of the preceding two types of methods, development history of deep learning and a brief introduction of involved works the essay has proposed.(2) Detailed and in-depth introduction for relevant visual sub-regions in the human visual cortex especially related visual processing for general visual features, detailed introduction for the four types of conventional ways to constructing image features and analysis of respective pros and cons. We propose to apply hierarchical learning pattern to acquire the image features through summarization on requisite characteristics of image semantic features on the basis of the preceding two procedures.(3) Introduction about distributed representation and clarification of the necessity for hierarchical representation in terms of the representation capacity. A brief introduction for the theory and learning procedures of general deep leaning models are also given.(4) Definition for image semantic classification issues. Meanwhile, in response to existing actual problems concerning image semantic classification, we propose to extract semantic features from the images by using stacked de-noising autoencoder and convolutional deep Boltzmann machine. And then we give the implementation of image classification based on the two models. We summarize on common techniques for optimizing deep learning models in response to the difficulties incurred for optimization for the deep learning model.(5) Verification of the effects produced on implementation about the task of image semantic classification by applying the two deep leaning models on the CIFAR-10and STL-10datasets. Summarization and predication on the prospect of image semantic classification are made based on analyses for the classification results.