Image Stabilization Based On SVM Motion Prediction

High-resolution Earth remote sensing satellite is important for national economic construction and national defense, and has always been the national focus of key technologies. While platform vibration is a major factor restricting high-resolution imaging, motion compensation is a key technology to solve this problem. For the reason that the traditional image stabilization has certain delay for the compensation, the paper proposes to utilize motion prediction in image stabilization, and focus on the research of auto-adaptive motion prediction.Firstly, we introduce the background and significance of the research, and summary the research status. We mainly study the two machine learning algorithms:artificial neural network (ANN) and support vector machine (SVM). Results of simulated tests show that SVM performs better than ANN in computing rate, accuracy and stability. And robust ability of SVM is verified fatherly by simulated tests, which means SVM is fairly fit for motion prediction of image stabilization system..Since SVM performance greatly depends on SVM parameters, finding the optimal parameters is quite necessary. We introduce the SVM parameters and three main algorithms for parameter optimization including grid search, genetic algorithm and linear search. After that, we analyze the behavior of parameter pair (penalty parameter C and the kernel width y), and sum up a rule that a rough line with high accurate exists in'good area'. According to the rule novel linear search is proposed, which is to search the optimal parameter pair along the rough line. Finally, high efficiency, high accuracy and good stability of novel linear search method are illustrated both from theory analysis and simulated tests.At last, we applied this method to real-time image stabilization system.1HZ-80HZ vibration models based on absolute vibration pixels are built and tested. How light intensity, exposure time and sampling frequency influent the stabilization performance are explained. GMG operator and vibration residual are used to evluate the compensation rate. The good results show that auto-adaptive support vector machine based novel linear search is feasible in image stabilization.