Synthesized images for pattern recognition

Abstract

Since there is no generic procedure for machine pattern recognition due to its complexity, ad hoc computer algorithms have been developed for each class of problems. In visual pattern recognition, depending on the area of investigation, it is difficult to obtain test images with the desired characteristics. Capturing the original images in the first place may require special and/or expensive equipment, setups, timing, lighting conditions, relocation of equipment and personnel. Due to these difficulties, researchers often reuse the same few available test images, which may compromise the thoroughness of the investigation. Also, when existing images need major changes in optical parameters (viewpoint, illumination, relative position of objects) the original environment and objects may not be available. This paper proposes generating graphic images by computer with characteristics close enough to reality for testing new pattern recognition algorithms. For demonstration of the idea, we use the graphics resources: ray tracing, fractals, texture-mapping and smooth surfaces, such as spheres and Bézier patches. Two illustrative examples are presented: (1) Moiré patterns by placing a cylinder array over a Bézier patch and sphere; (2) pavement distress images (cracks and potholes) from fractal backbone lines. At a relatively low cost, the image is synthesized in the form of a light intensity map file, which can be readily input to the pattern recognition algorithms under investigation. After an algorithm produces acceptable results using synthesized images, then we can validate it by capturing real images under the combination of parameters chosen in the previous step. In order to reduce costs, the field setup parameters may be decided from the best simulation settings. The proposed approach to the development of algorithms for pattern recognition will save time, money and resources at research sites where quality graphics can be rendered.