Study On Object Detection Model Using One-Stage Mode

The popularization of high-definition cameras visibly strengthen our ability to capture the meaningful scenes in daily life.At the same time,the massive data it collects in this procedure greatly fertilizes the science of computer vision.Thanks to the occurrence of high quality images and high-performance computing devices,object detection is able to utilize neural network to speed up its computation and gets better and better results.An object detection model can coarsely be cut into two parts,a feature extraction part to get the important information for recognizing as well as a classifier and localizer to make the final decision.It plays an important role in many popular intelligent applications,such as intelligent monitoring,intelligent driving and smart home.However,many existing models,such as RCNN,YOLO and SSD,struggle in the tradeoff between detection speed and accuracy,which are both important in real life.In this situation,we need to explore a better detection model which is more accurate in real-time premise.Specifically,the major works of this thesis can be summarized as follows:1.Focusing on time complexity and accuracy,we analyze the merits and demerits of existing detection models.Under the consideration of detecting speed or precision,different models pay much attention to the design on feature extraction network,detection procedure,data augmentation and loss function.We summarize the relationship between the effects and causes of RCNN,YOLO and SSD to give a good guidance to those who want to take part in this field.2.We redesign the feature extraction network of the detection model with dense connection,carefully pick or modify other excellent parts of other first-class detection model to make up our model,called DCOD.Based on fast one-stage detection model,we deepen the feature extraction network substantially to improve its ability of expression,thus improve accuracy.At the same time,the introduction of dense connection helps cut down extra calculations which usually are the side effects of deepening network.Finally,we utilize a slightly-modified loss function and a well-chosen training method to stabilize the training process.Experiments conducted on PASCAL VOC2007+2012 show that DCOD gets satisfactory result with 74.8 mAP and 53.7 FPS,which is comparable with other popular models,even if they have done more well-directed optimizations that we have not.3.Small object detection bothers many detectors.We explore the feature fusion method and propose a new fusion method to cope with this challenge.Shallow features contain more details.YOLOv2 reorganizes the shallow features into smaller features to help prediction,which is time-consuming and has semantic contusion.We replace the reorganize operation with a basic pooling operation,simplify the calculation and remove the semantic confusion.DCOD gets 75.6 mAP and 52.7 FPS after utilizing this opitimization,gets the best accuracy in real-time models.