Research On Paper Edge Detection And Extraction

Human-computer interaction technology in desktop augmented reality is widely used in education,entertainment and other fields,and is the key research direction of HCI.Paper is an important medium and main processing object in desktop augmented reality interactive system.It has the characteristics of weak texture and strong structure.In addition,there is interference from lighting,other targets and background in the real desktop interactive scene.Therefore,how to detect paper quickly and accurately is a basic and challenging task in a desktop augmented reality interactive system.This paper mainly studies paper edge detection based on full convolutional neural network and paper extraction.The main work includes the following aspects:?.A multi-sample paper edge dataset,named as MPDS,is built.Aiming at the problems of abundant target edges in the existing edge detection datasets and insufficient pertinence to paper edge detection,a paper dataset including 1500 samples is built in this paper.The pictures of A4 paper,letter paper and colored paper are taken in a variety of different desktop environments,such as ordinary backgrounds,complex backgrounds,interference objects,nonuniform lighting,etc.A semi-automatic method combined with manually labeling key points and programming is designed to mark sample to obtain the groundtruth of each samples.The MPDS provides data guarantee for the follow-up research of this paper.?.An efficient feature extraction edge detection network EFE is proposed.In view of the poor detection effect and robustness of traditional image processing edge detection methods,and slow detection speed of HED models based on full convolutional neural networks,EFE model is proposed.The MobileNetV2 architecture is used as the backbone of the HED network,the last two bottleneck blocks with large number of parameters,convolution layer with large number of output channels and pooling layer in the network are removed.Then an efficient channel attention mechanism is introduced on the backbone network,and a rich deep side supervision module and an upsampling fusion module are added.The experimental results show that the EFE model in this paper can greatly eliminate the interference of noise and has strong robustness.The model detection accuracy measures ODS and OIS are both above 0.9,and the detection speed is 39.02FPS,which achieves a balance between detection accuracy and speed.?.An edge thinning method with improved non-maximum suppression and a paper extraction method based on paper structure constraints are proposed.In order to avoid the problems of cracks,burrs and double edges during edge thinning,the non-maximum suppression method is improved,using linear interpolation to obtain values of pixels which locate in the gradient direction,and judging the maximum value of the gray value,and retaining pixels which gray value is greater than 250.The experimental results show that the method in this paper can achieve better edge thinning effect.In addition,taking advantage of the structural characteristics of the paper shape in the real environment,selecting conditions are designed,the missing vertexes are recovered when the paper is occluded by calculating the intersection of the edge lines.The experimental results show that the paper extraction method designed in this paper can obtain excellent paper extraction results in various complex scenarios.