Research On The Techniques Of Object Detection In Natural Scenes

Object detection,that is,to distinguish the target objects from the background and locate every instance given in an image,is an important and essential problem in computer vision.With the rapid development of Deep Learning,research in computer vision has come to a new phase,and object detection is gaining more attention than ever.Many applications such as self-driving cars,personal AI assistant and intelligent monitoring significantly rely on accurate and efficient object detection algorithms.In this paper we propose our research on three different aspects to improve deep-neural-network-based object detection in natural scenes,Firstly,for 2D pattern detection like logos and letters,which have relatively fixed shape,in order to tackle the challenge that Deep Learning need massive training data,we propose a method to synthesize training data.With few template pattern images,the program is able to generate mass training data for detection by simulating multiple effects of natural scenes via geometrical and texture transformation.Experiments show that similar performance is achieved using synthesized data compared with using hand-labeled data.This method dramatically reduces dependence on hand-labeled data and lowers the cost.Secondly,while Faster RCNN is a popular object detection neural network,it has difficult to detect small objects.We improve Faster RCNN by utilizing multi-scale features to generate region proposals and classify them.By concatenating feature maps of multiple layers,the network is able to extract high level semantic information as well as low level features,and thus has a better accuracy.Experiments on Pascal VOC show that our network outperforms ordinary convolutional networks that only use one group of feature maps on small object detection.Lastly,we implement a GUI system for training and deploying object detection networks.The system has a B/S architecture and is highly modular.It covers the complete training process,including dataset preprocessing,model training and model evaluating.With this system,non-professional users can train and deploy a deep neural network for object detection easily via a browser,which will accelerate research and application on object detection.