Subspace Construction from Artificially Generated Images for Traffic Sign Recognition
Recognition technologies using digital cameras have gained considerable interest in recent years. However, even with the improvements of digital cameras, the quality of captured images often can be insufficient for the recognition in many practical cases.
- PDF / 999,361 Bytes
- 22 Pages / 439.363 x 666.131 pts Page_size
- 97 Downloads / 188 Views
Subspace Construction from Artificially Generated Images for Traffic Sign Recognition Hiroyuki Ishida, Ichiro Ide, and Hiroshi Murase
Abstract. Recognition technologies using digital cameras have gained considerable interest in recent years. However, even with the improvements of digital cameras, the quality of captured images often can be insufficient for the recognition in many practical cases. In order to recognize low-quality images, similarly degraded images should be used for training classifiers. This chapter presents a training method for the subspace method. It is named “Generative learning method,” since the training images are generated artificially from an original image. Conventional approaches used camera-captured images as training data, which required exhaustive collection of captured samples. The generative learning method, instead, allows to obtain these training images based on a small set of actual images. Since the training images need to be generated on the basis of actual degradation characteristics, the estimation step of the degradation characteristics is introduced. This framework is applied to traffic sign recognition that is one of the important tasks for driver support systems.
1 Introduction to the Generative Learning High-resolution digital cameras have come into widespread use in recent years. Recognition technologies using such digital equipments are especially of practical concern. However, objects in distant places still tend to be captured in low-resolution and blurred, which has a serious effect on the recognition performance. The generative learning method [1] is developed to solve the problem of degradation. It was originally proposed for the recognition of hand-written characters [2, 3]. Hiroyuki Ishida Toyota Central R&D Labs. Inc., 41-1, Yokomichi, Nagakute, Aichi, Japan e-mail: [email protected] Ichiro Ide · Hiroshi Murase Nagoya University, Furo-cho, Chikusa-ku, Nagoya, Japan e-mail: {murase,ide}@is.nagoya-u.ac.jp
Y.-W. Chen and L.C. Jain (eds.), Subspace Methods for Pattern Recognition in Intelligent Environment, Studies in Computational Intelligence 552, c Springer-Verlag Berlin Heidelberg 2014 DOI: 10.1007/978-3-642-54851-2_4,
83
84
H. Ishida, I. Ide, and H. Murase
It generates artificially degraded samples, and allows to make classifiers trained by them. Traditionally, training images ought to be collected from actual images taken in the real world. Such a collection-based approach may be the most straightforward approach to obtain a set of training samples. In many practical cases, however, camera-based collection of a sufficient number of training images in various conditions is unrealistic. Let us consider collecting the training images for many categories. In the character recognition task [4], for instance, the number of categories tends to be large, and at the same time, printed text may even contain various types of fonts. This diversity of characters makes the collection difficult. Moreover, various conditions that cause respective distorti
Data Loading...
