Object Detection Based On Exemplar-ldas Ensemble And Multiple Instance Learning

Object detection contains a series of processes.Firstly,extract low-level features from static images or dynamic videos.Then,encode and combine these features.Next,use machine learning methods to obtain object detection results for the original images or videos.As the key point from the shallow image processing to the deep understanding of images,object detection technology has become an important research direction of the computer intelligence field,and has been widely applied to people's production and lives,such as the assembly robots based on machine vision in the industrial production field,military applications of the remote sensing aerial image processing,the video surveillance equipment applied to intelligence transportation,the fingerprint decoding and auto-focus cameras used in our life and so on.These applications of object detection technology not only lay a foundation for the modern economy,national defense,science and technology development,but also greatly improve the quality of people's life.Currently,the object detection technology still has a lot of room for improvement,such as how to reduce the training cost during the machine learning stage and maintain the detection performance at the same time,how to enrich the visualization representation of object detection results,how to improve objects learning and detection achievements under unbalanced data,etc.Aimed at the above problems encountered in the object detection technology,and inspired by the Ensemble of Exemplar-SVMs model as well as the Exemplar-LDAs object detection framework,this paper proposes an object detection model,which is based on Exemplar-LDAs ensemble and Multiple Instance Learning Algorithm,and achieves better detection performance.For the HOG features extracted from every exemplar,we train a corresponding discriminative Exemplar-LDA classifier with less training costs.According to detection results from the validation image set,we build a correlation matrix and a co-occurrence matrix to integrate the generated exemplar classifiers and adjust scores of detection windows.In addition,we construct a second layer of positive and negative sample packages with detection results from the validation image sets and the negative image sets.Furthermore,we use Multiple Instance Learning algorithm to train them and give each sample an accurate label,in order to construct a single category mi-SVM classifier.We apply the single category classification to filter out detection windows on testing samples which are found by the integrated exemplar classifiers.And it further improves the detection accuracy and reduces the false detection rate.Above all,exemplar models can transfer the available meta-data(segmentation,geometric structure,etc.)directly onto associated detection targets,which makes the description of target properties more accurate and richer.The object detection model based on Exemplar-LDAs ensemble and Multiple Instance Learning Algorithm,is implemented by the hybrid programming of MATLAB and C++.It is tested and analyzed on three categories of train,car and sofa from the PASCAL VOC 2007 database.According to the experimental results,compared with the original Ensemble of Exemplar-SVMs object detection framework,our model increases average precisions of 7.6%,26.4%,25.2% for the train,car and sofa respectively.It reduces the negative samples mining costs during the single exemplar classifier training process and enriches the detection results intelligent expression of exemplars at the same time.