Object Detection Based On Convolutional Neural Network

Object detection is aimed at locating and classifying all objects which is defined in an image.It is a fundamental task in computer vision,and it plays an important role in video monitoring,industrial detection,automatic driving and other application fields.With the improvement of computing power and the establishment of large-scale datasets,deep learning methods have achieved a great success in image processing,speech recognition and other fields.Moreover,convolutional neural network(CNN)has been widely used in the field of computer vision because of its great ability of representation and insensitivity to the change of illumination,scale and deformation.Object detection algorithms based on CNN have replaced the algorithms which based on manually designed features,and it have became significant methods at present.In this thesis,the object detection algorithms based on CNN in recent years are summaried and classified.And for getting a better detection performance,the object detection algorithm which is imperfect is improved.The works achieved in this thesis are as follows:(1)In the stage of detection in Faster R-CNN,the semantic feature of region proposal is the only thing that needs to be considered by the detector.It may lead to a classification mistake if the semantic feature of region proposal has a poor quality.Therefore,when a region proposal is classified,the effect of class-related coexistence relationship which is the contextual information is considered for reducing the probability of this classification mistake.Experimental results on object detection datasets indicate that our method is effective in improving performance.(2)Faster R-CNN has a poor detection performance on MS COCO object detection dataset.By doing an analysis,it is clear that this poor performance is caused by the large amount of small objects.And the detector can not adapt to the great variation of scale.Therefore,Faster R-CNN improved by constructing fusion feature pyramid,multi-scale RPN(Region Proposals Network)and multi-scale training is proposed.Experimental results show that our optimized algorithm gets a 4.7% improvement in mean average performance(mAP).