Image Registration Method Research Based On DM642 Processing Platform

Image registration is an important field in image processing, it refers to finding an transform, which can match and align two or more images that are taken at different time, by different sensors(imaging devices) or under different circumstances(weather, illumination, position and angle). It has been widely used in areas such as remote sensing data analysis, computer vision and medical image processing, etc.Image registration can be divided into two groups: the first group is based on area, such as template matching method; the second is based on characteristic point, such as corners. In this paper, both these methods are studied.1. Registration method for translation. Translation is a special kind of rigid body transform, its transform matrix has two variables. This registration method is an area-based method, so it doesn't has a step for characteristic detection, it processes images directly, the traditional method in this group either use the translation property of Fourier transform or use template matching algorithm, either way they process the entire image area, so they are all inefficient in certain way. In this paper, we propose a new method, which improve the traditional methods. It had two steps, firstly, a foreground extraction based on aprior knowledge is performed to located the foreground area that we are interested in. secondly, a template matching method combined with template checking method to find the best matching point in the search area. The template checking method aims at finding the best template from the randomly selected template, so it can significantly improve the speed of template searching process as well as the the accuracy of result compare with the traditional template searching method.2. Registration method for scale transform and affine transform. This method is based on characteristic point, for example corners. Harris corner detector is a classic tool to extract feature, it is stable to illumination change and rotation but unstable to more complicated transform. In order to register images with different viewpoint, we extend Harris corner detector to scale-space to gain invariance to scale change, then we apply affine shape adaptation to the scale invariant point until convergence is reached, giving it invariance to affine transform. With these local features, we use general feature descriptor and matching algorithm to generate matches, then we use the matches to calculate the geometric transform matrix, which enable the final registration. Result shows our algorithm can get accurate matches than SIFT, and less difference between registered images.Recently, with the development of micro-electronics technology, the multi-media technology is widely used in every aspect of our everyday life, especially the embedded software development of the image and video acquisition and processing. The Digital Signal Processor has greatly helped the development of high speed image and video processing system, the TMS320DM642 digital signal processor is especially designed for video and image processing by Texas Instrument, it has great computing power and many different kinds of peripherals. In this paper, we combine image processing with DSP technology and designed an image processing platform. The platform had two kinds of input, infrared light camera and ordinary camera, the core of the platform is an evaluation board based on TMS320DM642 DSP and many other peripherals, a LCD television is used to display the output signal. All of these devices are connected together to form an image processing hardware platform. Based on this platform, we implement the above two algorithms,1. Monocular stabilized object tracking system. Infrared light has its own characteristic, infrared camera is often used military field, it's often mounted on mobile platform such as plane or satellite. Consider the long range sensibility of infrared camera, it often work with long-focus lenses, hence, it must prone to vibrate. In order to make later process possible, an image stabilization step must be taken. We consider the common translation caused vibration, use foreground extraction and template matching to achieve stabilization, then we use template matching and other image preparation technique to calculate the frame-to-frame offset and use the offset to control the tilt-pan camera. In this way, we allow the camera to center on the object.2. Binocular image registration system. Binocular vision is also called 3-d vision, it emulate the human vision system, the basic idea is to observe one scene from two perspective, and use triangulation to calculate the difference between pixels in order to get 3-d information of the scene. This process is the same with human eyes. When we acquire two image from different angles, the images is bound to deform from one another, consider the case of affine transform, we applied the improved Harris corner extractor to get corners that are stable under affine transform and finally achieve registration, proving the effectiveness of this method.