Studies in Human Vision Model Applied to Image Recognition

[[abstract]]本論文對人眼視覺模型及其辨識特性提出一個新的定義，並以此為基礎導入了新的人眼辨識神經網路方法，首先以調變傳遞函數(MTF)為基礎對所有的人眼視覺模型辨識能力加以研究，再以新方法將最適合人眼的視覺模型加以形狀因子修正，論文中廣泛探討了包括人眼像差及人眼光學模型，人眼的辨識特性及各種人眼視覺模型的驗證。 雖然光學及電子技術的進步使圖像的雜訊減少許多，但人眼對低解析度或低訊雜比的圖像辨識能力仍然有困難，由於生理、心理、像差及疲勞等因素，限制了人眼辨識的能力，此外，因顯示器上亮度變化，鏡頭雜訊等因素，對人眼辨識能力上亦很大的干擾，這篇研究找到了適合在遠距離及軍事上用途的辨識方法，論文中提出了以倒傳遞神經網路(BPNN)及形狀因子修正方法來改進人眼的辨識能力，實驗結果顯示可以穩定的提昇辨識能力到95%以上。[[abstract]]This research proposes a new definition of human eyes vision especially for human eyes model and recognition MTF characteristics. Based on theoretical definition derived, this research will introduce newly developed method for enhancement of human recognition via neural network method. With regard to new theoretical definition, firstly human recognition capability are well-evaluated for all the studied models by modulation transfer function (MTF) curve, later an newly developed human recognition model with correct factor will be well-defined. Those theoretical aspects that have been explored in this research have focused on the following: human eye modeling, including aberration analysis, creative human eye optical model, human vision recognition MTF characteristics and various human vision models verify. With regard to human recognition enhancement by neural network, there are still some difficulties although modern optics and electronics has their decent resolution and ratio of signal to noise. In addition, observers have its limitations with regard to psychological and physiologic fatigue and aberration as well, which might complicate recognition work. Besides, some compromises obviously arise from the noise inherent in the lens geometries, brightness and darkness on the display. It will be critical in this research to find the proper and convinced auxiliary method to guarantee the best accuracy of recognition especially for military, remote sensing and arsenal purposes. In this thesis, back-propagation neural network (BPNN) is proposed to auxiliary strengthen capability of human recognition. Experimental results show that human recognition will be stably improved to above 95% despite of shaping factors applied in this research