Dissertation
Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

Depth Estimation from Monocular Infrared Images Based on Kernel Principal Component Analysis and Neural Network

Author LiLinNa
Tutor SunShaoZuo
School Donghua University
Course Power Electronics and Power Drives
Keywords Infrared image Depth estimation BP neural network KPCA
CLC TP391.41
Type Master's thesis
Year 2013
Downloads 110
Quotes 0
Download Dissertation

100years ago, the rise of an information exchange and processing technology based on infrared radiation. It is mainly to study the law of the emission, the transmission and the application of the principle. In recent years, there is a growing emphasis on the exploration of infrared technology. With infrared imaging technology to flourish, the role of infrared images have become increasingly prominent in the military and the civilian fields. People gradually improve the image quality requirements of the infrared image.Because of the special imaging mechanism which converts the temperature distribution of the object surface into a visible image of the adult eye in a special installation, a major flaw of the infrared image is the lack of a sense of depth and sense of space. To achieve a sense of space in an infrared image, it can greatly improve its visual quality, to improve understanding of the scene and is conducive to the image analysis and processing. To Solve this problem, This paper presents a depth Estimation from Monocular Infrared Images Based on Kernel Principal Component Analysis and Neural Network. Image depth estimation is calculating the distance information from the image. Essentially, that is a question of depth perception. At present, there are three methods of depth estimation:the depth estimation from binocular image,the depth estimation from image sequence and the depth estimation from monocular image Among them, the first two methods have found wide applications. And they rely on feature difference from two or more images.But the depth estimation from monocular image can extract depth clues according to prior knowledge and inherent characters. The infrared images in practical applications often is a monocular image. The paper will mention the method of depth estimation is monocular depth estimation method, with unify the infrared image inherent characteristic to infer depth as machine learning problem. The paper’s main work is the research of depth estimation from monocular infrared images. It includes:(1)extract depth features from infrared images;(2)screening of crude extract features based on KPCA;(3)with the neural network theory as the foundation, obtain estimation model by training selected depth features. The paper’s innovative points as follows:1. Based on KPCA algorithm of the linear kernel function, polynomial kernel function, Gaussian radial basis function, the Multilayer Perceptron kernel function, we can extract depth features from crude extract features. And using Particle Swarm Optimization, it can optimize the nuclear parameters of KPCA.2. Based on BP neural network which is the most active branch of machine learning research, we train depth features after screening in order to get the depth estimation model.Proved by experiments, kernel principal component analysis (KPCA) which does not require direct samples nonlinear mapping is simple and efficient. The method can not only guarantee the correlation between feature vector and depth value, but also ensure statistical independence from each feature vector and avoid the curse of dimensionality. The depth estimation model based on the polynomial kernel function can better recover the depth information of the original image. Compare depth estimation algorithm based on image segmentation, this paper presents a nonlinear estimation model based on Kernel Principal Component Analysis which has better generalization.

Related Dissertations
More Dissertations