EE368 Project: Visual Code Marker Detection through Geometric Feature Recognition

Visual code markers are developed as a humanmachine interactive tool [1] [2]. In this class project, a code recognition program is developed to read code markers. The requirements of the code recognition program are high accuracy, speed and illumination insensitivity. There can be more than one visual code in the picture and the orientations of the visual codes are not restricted. The code markers are taken with cell-phone cameras. Images from cell phones are poor due to several reasons. The images typically have a soft focus, caused by low quality lens, camera shake, and color sensor interpolation. Motion blur is exacerbated by the slow lens’ poor light gathering ability. A low signal to noise ratio is the result of cheap sensors, analog circuitry, and lack of incamera noise canceling hardware/software yielding significant Gaussian-like additive noise. Inadequate metering of the mobile phone camera results in poorly exposed photos. In addition, the camera sensor has low dynamic range preventing postprocessing algorithms to recover details, resulting in low contrast in images. Hence, cell phone photos are more challenging for the software to recognize. The most significant challenge might be the viewangle induced distortion. At first glace, the visual code detection problem resembles a license plate recognition problem. Various techniques have been presented for recognizing license plates based on edge detection and Hough transform [3] [4] [5] [6] [7]. Others have proposed hybrid methods combining color segmentation and statistical techniques for more accurate segmentation [8] [9]. Hough transform is a popular solution to image segmentation, as license plates typically appear square and upright in the pictures. Further, there is usually no more than one license plate per picture. Most algorithms extract vertical edges and find the most prominent vertical lines to isolate the license plates. However, visual code detection is not suitable for this technique. Visual code markers do not have closed and clear boarders, and they could have long edges in the data field as well. Furthermore, the visual code markers typically appear in highly complex environments such that the markers’ perimeters are not prominent compared to its surroundings. In this report, we present a visual code detection algorithm that combines edge detection, geometric feature recognition, and an error rejection metric to improve reliability. It does not include the Hough transform.