Article Preview
Top1. Introduction
With the continuous development of digital technology, a large number of digital images appear in people's life every day. Image is the description of the similarity and vitality of the objectives. It is also the most common information carrier used in the human social activity and one of the main information sources (Zhou, Feng, Ma, & Yuan, 2016). With the expanding of the Internet, the scale of the image data is increasing. Artificial recognition is not able to deal with the mass scale of image data. Therefore, computer image recognition becomes an important research topic currently. Image classification, as an important research method of computer image processing, is the research focus in the field of image processing in recent years.
The original image classification technology first annotates the image manually and then classify using the annotation information. In this phase the image classification was only based on the text annotation. The accuracy of this method relies on the subjective consciousness of the note annotator. It is replaced by content-based image classification (Mistry & Ingole, 2013; Dharani & Aroquiaraj, 2013; Yang, Yu, Gong, & Huang, 2009; Krizhevsky & Hinton, 2011) rapidly because of the high artificial cost and low efficiency.
Feature extraction is the key process of the content-based image classification. Among these SIFT, GIST and HOG are the most popular image feature descriptors. In reference (Lowe, 2004), SIFT features was used to identify the different images appearing in the same target. Translation, scaling and rotation were stable, but the main direction could be inaccurate because it relied on the local pixel gradient direction too much. The subsequent steps of feature vector extraction and matching were heavily dependent on the main direction and the small angular deviation could also cause the feature matching error. Both of these lead to the failure of the matching process. Reference (Zhao & Ngo, 2013) proposed a F-SIFT algorithm based on the rotation invariant features of SIFT algorithm and applied this method to the large-scale video copy and target recognition. GIST feature put forward by the reference (Oliva & Torralba, 2001) was used to extract the spatial structure characteristics which reflect the scene category and to ignore the subtle texture information of object or background. This design made it achieve better performance in the outdoor scene classification but not so satisfactory in indoor. Reference (Tan, Yang, & Ma, 2014; Dalal & Triggs, 2005) applied HOG feature to the human face detection and had a good outcome. Reference (Bristow & Lucey, 2014) used SVM to train current HOG feature in order to detect and classify the image. It received a nice result in the face detection and pedestrian detection tasks. Reference (Felzenszwalb, McAllester, & Ramanan, 2008; Forsyth, 2014) built DPM model base on HOG feature and applied the module to the scene classification task. This method took the advantage of the accurately detecting of DPM and enabled to increase the overall classification accuracy. But due to the poor stability, there was big classification accuracy difference between different categories.
These content-based algorithms mentioned above occupied the mainstream of image classification algorithm before the popularization of deep learning theory. But the success case was rare because the manual feature design required a lot of experience and well understanding to the related field and data. And a large amount of debugging tasks was required after the features design. A classifier algorithm was needed base on the manual feature design, which was another difficulty. It was nearly impossible to merge and get the optimal effect when performing the manual feature design and classifier algorithm design at the same time.