Fast Object Detection And Recognition Based On Several Samples

Object detection and recognition in images(or video frames) are the core problems of computer vision research. Although there have already existed many approaches of detecting and recognizing objects, these approaches can not be applied to real world very well. The problems are:(1) Lots of samples are needed to train a classifier, and to locate the object in a target image;(2) Sliding window as a high cost detecting strategy, is used to locate the object in a target image. In real world applications, it is difficult for people to find thousands of samples for an object. Usually, there are just few samples in hand(usually several or dozens of samples). In this case, it is impossible to train a classifer to recognize the objects in a target image. Besides, it is not necessary to use the time-consuming sliding-window strategy to achieve object locating. Based on the statements mentioned above, this work focuses on few sample application. A few approaches are proposed to organize samples efficiently and locate objects fast and efficiently.The contributions of this work are:(1) A local adaptive steering feature(LAS feature) is proposed to represent local feature of an image. Based on this LAS feature, a method of constructing voting space is proposed to express the tolerance of LAS feature at each location. In object recognition stage, the similarity between two images is computed through voting strategy pixel by pixel using the trained voting space. And the sliding window is used to detecting the object in target images based on LAS feature and voting space. In the experiments, this proposed method is better than the existed template matching methods, and also better than some detection methods which use lots of samples to train a classifier on recognition accuracy. So it is demonstrated the practical and effectiveness of the proposed method.(2) A fast object detection method is proposed based on several samples. A new detection strategy is adopted, which is based on image patches instead of sliding window. This new strategy reduces the computation complexity two orders of magnititude. Besides, each location of feature in voting space is assumed to subject to Gaussian distribution. Parameter estimation is used to compute the parameters of voting space. And the cell of voting space is achieved to be adaptive. The description of local feature is more effective when the voting space is used to detecting object. The experimental results demonstrate the proposed method achieves object detection fast and accurately.(3) A multiclass object detection method is proposed based on two-demension inverted index table constructed by two-demension features. Firstly, the two-demension feature is computed through local gradient intensity and orientation. Then, binary codes are obtained from the two-demension feature by encoding. And the binary codes are used to construct two demensional indexes. Each index is corresponding to a nubmer of voting locations with multiclass flags to achieve multiclass object detection. The experimental results demonstrate the proposed method achieves multiclass object detection accurately in real-time.(4) A foreground detection method based multiple auto-encoder networks is proposed. Firstly, when the background changes are not large, one auto-encoder network is constructed to extract background images which do not contain foreground images based on a few inputing video frames. Then these background images are as input for another auto-encoder network. The second auto-encoder network is used to learn background. When the background changes are large, a learning algorithm with adaptive background tolerance is proposed. Two auto-encoder networks are constructed through multi-hidden layer networks. And the tolerance of background is as the parameter of cost function of the first deep auto-encoder network to learn adaptive tolerance of dynamic background. Besides, an online learning method is also proposed based on deep auto-encoder network. By constructing the sensitive function of the cost function, the weight vectors of auto-encoder network is updated and substituted to achieve fast online learning. The experimental results prove the effectiveness.