Research On Learning Based Super-resolution For Millimeter Wave Image

The passive millimeter wave(PMMW) imaging system achieves imaging by measuring the millimeter wave energy difference of human body and concealed objects. The advantage of non-radiation and non-violation make the system very suitable for the security inspection of human body. However, because of the imaging mechanism, the imaging resolution of passive millimeter wave imaging system is limited by the antenna, which will cause geometric distortion and the loss of detail information of high frequency, leading to distortion and blurring of image. To solve this matter, dislocation needs to be corrected and the image super resolution algorithm is deployed to raise the resolution of PMMW image.Aiming at the dislocation between even-odd lines of PMMW image acquired by the single channel PMMW imaging system, the correction algorithm is researched. The super resolution algorithm is researched to raise the resolution of PMMW images. The major work of this dissertation include:1. The PMMW imaging theory is studied, and the mathematical model of imaging is analyzed.2. The single channel imaging process and the cause of dislocation between lines are analyzed. The dislocation estimation algorithm based on frequency domain is proposed to effectively correct the dislocation between lines.3. The initial dislocation estimation optimizing algorithm based on neighbour masking is researched. Meanwhile, the polynomial fitting dislocation estimation optimizing algorithm based on cross validation is proposed. By cross validation, the best fitting order is acquired to effectively solve the problem of over-fitting and under-fitting, and the false estimation as well as false correction of dislocation is reduced, providing important pre-processing base and results for super resolution processing.4. The super resolution theory based on learning and its core technique is researched. Using combined feature extraction, feature dimension reduction and discrete cosine transform based pre-classification, the miss matching and high computational problem in learning based super resolution is improved. The simulation results shows that the improved algorithm performs better in both subjective and objective matter, and has less running time consumption.5. Sparse representation theory is researched and its application in image super resolution is discussed. Using pre-classification, the conventional dictionary learning algorithm is improved, providing better matching performance between dictionary and samples. The simulation results shows that the improved algorithm has better imaging quality and less computational complexity.