Research On Key Technologies Of Object Detection Based On Region Proposals

As an important branch of image processing and computer vision,object detection has been widely used in civilian and military fields,including: automatic driving,intelligent monitoring,tumor screening,and precise guidance of weapons.The core task of object detection is to accurately recognize and locate the object in the image using the visual perception model and search strategy for a given image of any size.With the development of artificial neural networks,solving object detection problems using deep learning methods is still a research hotspot in the world today.In particular,the speed and accuracy advantages of the object detection framework based on convolutional neural networks have prompted more and more researchers to devote themselves to solve object detection problems using deep convolutional neural networks.Although the performance of the object detection algorithm based on convolutional neural network has been greatly improved,the accuracy of the object detector cannot meet the needs of the actual scene,especially the detection accuracy of the multi-scale objects,due to the deviation between the representing model of object and the human visual perception system.This dissertation mainly starts from the route of improving the quality of object region proposals and improving the performance of object detection framework,and conducts research on multi-scale object detection to achieve fast and accurate detection of multi-scale objects in visual scenes.The main work and innovations of this dissertation are as follows:For large and medium size objects,a region proposal algorithm based on richer convolutional features network and object saliency is proposed,which improves the recall of region proposals for large and medium size object.Firstly,the richer information of objects boundries are extracted using richer convolutional features networks;Then,the salient features of objects are described by the property of color contrast between object regions and their surroundings;Thirdly,the space position model for the objects in the scene is established;Finally,the region proposals are generated using the richer information of objects boundries,the salient features and space position model of objects.The experimental results on PASCAL VOC 2007 test set show that given 500 object region proposals and a fixed intersection ratio of 0.5,the recall of the proposed algorithm is 1.74% and 2.59% higher than selective search algorithm for large and medium size objects.At the same time,the proposed algorithm takes 0.76 seconds to process an image that contains different scale targets on a CPU@4.20 GHz computer,which improves the computing efficiency of the algorithm compared with the Select Search algorithm.For small size objects,a region proposal algorithm based on richer convolutional features network and superpixel saliency is proposed,which can improve the recall of the region proposals for small size objects.Firstly,the super pixels are obtained through performing segmentation algorithm on the image;Then,the saliency of the superpixel is defined according to the color difference between this superpixel and its adjacent superpixels,the spatial position and integrity of this superpixel.Finally,the region proposals are formed by edge features generated from deep neural network and the objects saliency.The experimental results on PASCAL VOC 2007 show that the proposed algorithm can improve region proposals quality of small size objects.At the same time,2000 region proposals generated from the proposed algorithm can obtain the highest mAP(mean Average Precision)value than other approaches in Fast RCNN network,and this can indicate the proposed algorithm has good performance.In view of the difficulty of dealing with small size objects in large scenes,the improved Fast RCNN and the developed SSD algorithm are proposed,which significantly improve the detection accuracy of small size objects.The quality of region proposals is improved by using the smaller scales of anchors and convolutional features from shallow layers in RPN network of Faster RCNN.The context information of objects is introduced to SSD network to improve the detection accuracy of small size objects.In particular,in order to verify the detection performance of the proposed algorithm on small size objects,VOC?MRA?0.58 data sets were produced,and the experimental results showed that the improved Faster RCNN algorithm has about 6% higher mean average precision(m AP)than Faster RCNN;the developed SSD algorithm raised m AP by 7% than SSD methods.