Neural Network Based Feature Point Detection Method For Perspective Optimization

Feature point detection method of digital image is an indispensable part in the field of computer vision.At present,image matching problems in computer vision applications are usually processed by traditional feature point detection methods.In one type of these matching problems,there is perspective transformation information between the images to be processed.With this information,we propose a new feature point detection method based on neural network for perspective optimization.Firstly,we add perspective transformation information into the training data to maximize the recognition and judgment ability of the model.Secondly,we add the probability output of the model and the position deviation of the point relative to the center point of the original image to obtain a rank value.Finally,we choose the better ones by local non-maximum suppression operation with their rank value.Experiment shows that this method outperforms traditional methods in detection repeatability and gets lower error in the visual odometry instance.In the computational performance,this method can run at a realtime frame rate.