On a Shape Adaptive Image Ray Transform Ah-Reum Oh School of Electronics and Computer Science University of Southampton Southampton, UK aro1g11@ecs.soton.ac.uk Mark S. Nixon School of Electronics and Computer Science University of Southampton Southampton, UK msn@ecs.soton.ac.uk Abstract— A conventional approach to image analysis is to perform separately feature extraction at a low level (such as edge detection) and follow this with high level feature extraction to determine structure (e.g. by collecting edge points using the Hough transform. The original image Ray Transform (IRT) demonstrated capability to extract structures at a low level. Here we extend the IRT to add shape specificity that makes it select specific shapes rather than just edges; the new capability is achieved by addition of a single parameter that controls which shape is selected by the extended IRT. The extended approach can then perform low-and high-level feature extraction simultaneously. We show how the IRT process can be extended to focus on chosen shapes such as lines and circles. We confirm the new capability by application of conventional methods for exact shape location. We analyze performance with images from the Caltech-256 dataset and show that the new approach can indeed select chosen shapes. Further research could capitalize on the new extraction ability to extend descriptive capability. Keywords- computer vision, feature extraction, shape extraction, Image Ray Transform I. INTRODUCTION The deployment of computer vision has become more important in an increasingly industrialised society, requiring sophisticated methods for feature extraction in image interpretation. The conventional approach in computer vision is to perform feature extraction at a low level, to pre-process an entire image. This is followed by high level feature extraction to determine structure. Physical analogies have been deployed previously times for low level feature extraction. For example, [1] refined anisotropic diffusion [2] to find moving edges using heat flow in the temporal dimension. Force [3] and magnetic [4] fields have also been developed for image segmentation. Apart from these methods, there is a variety of feature extraction techniques using physical properties, such as time [5]. These operators act at low- level, prior to image segmentation. The Image Ray Transform (IRT) was originally a low-level operator which operates by analogy to light and has been applied as a pre-processing stage for image analysis within ear biometrics and retinal structures [6,7]. The IRT uses an analogy to reflection and refraction of light rays, with materials’ refractive indices calculated from the image information. The analogy actually guides development of the image processing operator; an exact implementation of light modeling is not used. The original IRT is a powerful technique for low level feature extraction and appeared especially suited to the detection of curved objects, as shown in its deployment in detecting ears for biometric purposes. Our novel extension to the IRT adds a new parameter, a shape factor, to allow edge detection, and object selection at the same time. This is then the first approach to combine low-low level analysis with high level structure. This is illustrated in Fig. 1 (a) which shows a synthetic image of a square and a circle. The original IRT provides the result in Fig.1 (b), showing the edges of the square and of the circle – and the result of applying the Hough transform is to find neither the line nor the circle, precisely. By introducing the new shape factor, the process can be arranged to select circles, as in Fig. 1 (c) – where the circle is correctly detected, or lines, as in Fig. 1 (d) where a line is detected. This is achieved by changing the value of a single parameter – the shape factor. The Hough Transform [8] can be used to corroborate the quality of the new feature extraction process. Accordingly, adding the shape parameter is a new way to extract features of objects and extends the whole performance of the IRT. Application of the Hough Transform shows that the extended IRT can lead to more general capability, with superior performance, to that of the original IRT. (a) original image (b) original IRT (c) extended IRT shape factor = 1.5 (d) extended IRT shape factor = 1.9 Figure 1. Illustrating results of the original Image Ray Transform (IRT) and the extended IRT, with the Hough transform results superimposed showing that the circle and the line have been emphasized by different selection of the new shape factor.