Research On Object Detection Based On Multi-instance Models

With the in-depth study of object detection, its development shows a trend fromhard to easy. At the beginning, researchers were trying to extract features fromimages which are invariant to image scale and rotation, and they want to extractfeatures which can absolutely represent the object. Research in this area has madesome achievements as the raised of some invariant features. There are manymethods to match objects and classify images, but it’s still hard for us to detect theobject in images.Find it hard, and then researches begin to deal with more simple questions.They just detect human faces and pedestrians from images. As the proposal of HOGfeature, the problem seems to become simpler. The basic idea of HOG feature is thatlocal object appearance and shape can often be characterized rather well by thedistribution of local intensity gradients or edge directions, even without preciseknowledge of the corresponding gradient or edge positions, so it’s easy to detectsingle shape object. As the common method always attempt to represent an objectcategory with a monolithic model, or pre-defining a reduced set of aspects, it’s hardto detect the true expression of the target, so the test result is not good. Then themulti-component approach for object detection has been raised. The main idea ofthis method is to training a separate linear SVM classifier for every exemplar in thetraining set. This paper is based on this method, and the main work of this paper isas follows:1. Analyzing the multi-component method for object detection form the size ofthe training data sets, this contains20categories of detection object. Experimentsshow that the detection performances vary insatiably as the variation of data set todifferent categories. The main reason is because of the diversity of the training data,as we can see each train sample will have a certain number of test results, but thedetection performances changes over the match of the training data set and test dataset under the HOG feature, therefore, it shows instability. So, we will analysis thelimitations of HOG feature and the construction of the template library to improvethe performances of the system.2. Taking into account of the limitations of the HOG feature, we tried tocombine SURF feature with it to improve system performance. As there areuncertain numbers of SURF key points, we will use SURF in calibration step.Combining the initial detection scores and the SURF match between the detectionresult and the train template, and the total scores were used to generateco-occurrence matrix, so as to compensate the limitations of both detection features. 3. Find the best way to construct the template library to improve theperformances of the system. First we analysis effective method to enhance thesystem recall, the main algorithm is a k-means clustering algorithm to the trainingdata set and a clustering algorithm based on the classifier trained in the use of SVMof a certain picture. By this way, we can improve the performance of the detectionsystem, and the principle of this method is to invade a certain category into anumber of parts, which can maximize the use of the ability of HOG feature.Experimental results show that this method not only improves the detectionperformance of the system but improves the testing speed also, and we finallycomplete the object detection system based on multi-instance models.