Image Classification Based On Deep Learning Models

The basic idea of image classification is automatically categorize all kinds of images with a computer through a variety of machine learning algorithms. Affected by irregularity, scale and light in image, there are great challenges to image classification. The recently developed deep learning method has an outstanding performance in the image feature learning.Compared to some shallow machine learning models, the deep learning model has a stronger expression and generalization ability for complex classification problems, especially when the target object has rich information. The model structure through layers of nonlinear network structure to represent the internal distribution of data, reflecting the nature of the input sample data.This paper explores some deep learning model, and focuses on feature learning in image classification with deep learning model.This paper’s main works are as follows:(1) A hierarchical feature representation combined with image saliency is proposed based on the theory of visual saliency and deep learning. The image sparse representation is introduced in the level of salient information, which compresses the feature representation and strengthens the semantic information of the image. Instead of using hand-crafted descriptors, our model learns an effective image representation directly from images in an unsupervised data-driven manner. The experimental results on two commonly used benchmark data sets Caltech 101 and Caltech 256 show that our method significantly improves the performance compared to those with the single sparse coding using local features.(2) This paper also explores the application in scene classification with convolutional deep belief networks model, which combines the advantages of convolutional networks and deep belief networks. The convolutional network has a good adaptability to changes of position, zoom, and rotation in image, but it ignores the high-order statistical features in the image. However, the DBN model has a good performance in extracting higher-order features in the image. Combining these two advantages, we apply it to the scene image classification and get good performance in testing data set.