Study On Non-uniformity Correction Of Infrared Image Based On Computational Intelligence Algorithm

As the key component of modern infrared imaging system, each unit of theinfrared focal plane array response inconsistently to the same radiation, which led to thegeneration of non-uniformity, by the performance of various noise of differentfrequency on the image, and has thus become the most important reason to curb of theinfrared imaging system to output clearly identifiable image. The correction ofnon-uniformity is necessary in actual applications, regardless of the visible requirementsof image, or the needs of subsequent processing.Since the infrared image contained non-uniformity is difficult to obtain, the thesisfirst analyzed noise characteristics of the infrared image as the basis, then simulate thenon-uniformity by adding random noise to the image, but also describe thenon-uniformity influence differently on the image quality by adjusted the noiseparameters.The main job of the thesis is to study on a typical kind of the scene-basedcorrection methods-the neural network correction method, and which is combined withthe computational intelligence algorithm. After the research of the classic neuralnetwork non-uniformity correction algorithm, we found that the selection of the hiddenlayer algorithm and the determination of initial parameters in correction impact greatlyon the convergence speed of the algorithm and the final correction effect.In the traditional neural network correction method, the hidden layer calculatesarithmetic average of the four neighbors of the center pixel domain as the expectedvalue to update the calibration parameters, which is simple and easy to calculate, but atthe same time, it ignores the different influence of the neighboring pixels to the centerpixel. Therefore the following improvements are proposed:(1) The window size was selected, and the arithmetic average of all the pixelsaround the center pixel in the window was calculated as the expected output;(2) The weighted average of neighborhood pixels within the window wascalculated, wherein the weight value is obtained in accordance with the gray-scaledifference between the neighborhood pixel and the center pixel; (3) Within the selected window, the membership function in the fuzzy logicdomain was used to classify the neighborhood pixels of center pixel, then each weightvalue was determined by the difference of area attribute, and the weighted average wascalculated as the expected output.The initial correcting parameters were usually determined by the initial randomdecimals or the prior knowledge. As an intelligent optimization method, particle swarmoptimization algorithm could be used to train the neural network on the first frame inorder to obtain a set of optimal values, and then the other frames could be corrected, sothat the convergence rate could be enhanced.After analysis of the simulation results, the improved algorithms proposed in thethesis had either achieved a better correction results, or improved the convergence rate,or decreased the complexity of the algorithm.