Object Detection Based On Convolutional Neural Network And Context Model

Object detection task is to identify each object in the image, and given their position. the classic object detection algorithm is DPM(Deformable Parts Model) algorithm, it uses a sliding window method to generate the candidate window, and then use the HOG(Histogrrams of Oriented Gradients) features and linear SVM(Support Vector Machine) classifier to classify the candidate window. The disadvantage of the DPM algorithm is computationally intensive and shallow HOG features. In 2014,Ross Girshick proposed R-CNN(Region Convolutional Neural Networks) algorithm to detection object, it use the selective search algorithm to generate candidate window, and then apply convolutional neural network on each candidate window to extract the feature of image, and then use the linear SVM to classify the candidate window. RCNN algorithm in PASCAL VOC datasets achieved significantly higher correct rate than the DPM algorithms.This paper is based on R-CNN algorithm, R-CNN algorithm introduces a selective search algorithm in preprocessing stage, opened up a new framework for object detection, namely the "selective search + convolutional neural network." However, RCNN algorithm in the post-processing stage using the traditional NMS(non-maxima suppression) algorithm. NMS algorithms exist two drawbacks, on the one hand, how to select the appropriate threshold is a difficult matter. On the other hand, NMS algorithm does not consider coexistence with the spatial relationship between objects in an image.This paper presents object detection algorithm combined context and R-CNN algorithm. We introduce a model of learning context, to characterize the spatial position exists between the various types of objects(which may be of the same class may be different categories), and given their learning and inference process in detail. When you choose the best candidate window, use it instead of the NMS algorithm. We have been training and testing this algorithm on the PASCAL VOC 2011 dataset, and compared with R-CNN algorithm which used NMS algorithm. Experimental results shows for often occur simultaneously in one image and have a particular spatial relationship between the objects, this method has significantly improved than R-CNN algorithm in accuracy.