Research On Image Object Detection Based On B-HOG

Using the computer to think as human to detect the contents of an image is an important research area in computer vision. Whether the output is a specific object divides the task into object detection and object classification which are the basis of image segmentation, feature matching, target tracking and other more high-level visual research and it has been widely used in face recognition and so on.Image object detection includes the following steps: firstly, scanning sliding windows on the image map to acquire feature points or feature vectors for each window; secondly, selecting the feature to classify; then to fuse the detection results and delete some redundant windows to the final result to give out the outer rectangle frame of the object area. Since it needs a lot of feature point extraction and matching tasks, the BING and HOG still could not fulfill the speed and accuracy requirement in some real-time research, especially could not be widely used in security environment.According to this, this paper proposes a new method based on binarized histogram of oriented gradient(HOG) and Ada Boost algorithm combined with the binarized technology in recent years. It could improve the detection accuracy and speed at the same time. First of all, we normalized all detection images into five fixed scales and used the high 4 bits of the gradient pixel to approximate the original one which would reduce the calculation amount. Considering the way that internal data storages, we use linear basis vectors to binarized these HOG features which would be converted it into fast operation and bitwise. And thirdly, the new size of HOG detection window and the new division of gradient directions would lead the block to be storaged in one byte. Considering that linear SVM classifier has shortcomings with difficult samples, this paper uses the Ada Boost algorithm and analysis quantitatively about its false detection rate, recall rate, generalization ability and the number of the cascade classifier. We also set in the the number of weak classifier in the training process to make sure that each level of the detection is not infinite,each strong classifier could classified more efficienct.In this paper, we use VOC PASCAL 2007 database for training and testing and then compared the results to some state-of art algorithms like BING, HOG and SVM, HOG and Ada Boost. The results show that our method has a certain improvement in detection time and detection accuracy.