An adjustable contrast optical target equipment was constructed. After researching the rela-tionship between image contrast and optical contrast, a contrast calibration method by the improved Back Propagation(BP) neural network was proposed. Firstly, the BP neural network model was designed for calibrating the contrast. Then, by combining the Levenberg-Marquardt(LM) with Shrink-ing-Magnifying Approach, the BP neural network was improved to optimize the convergence speed and generalization ability. Finally, based on the experimental platform of the adjustable-contrast target, the image contrast was obtained by measured radiation data. Comparing with the traditional BP algorithm, the improved one has a better convergence speed and generalization ability. Its calibration accuracy has been improved by 100 times and by 10 times as compared with those of the traditional BP network and the steepest descent method, respectively. When the training times is to be only 2 876 times, the maximum error between calibration value and target calibration value for the contrast is 0.01%, the training mean square error converges is 0. 000 459 441, and the test error converges is 0. 000 467 003. These results demonstrate that the algorithm is feasible and can meet the demands for contrast calibration in the equipment.%搭建了可调对比度目标源装置,研究了图像对比度和光学对比度的关系,提出了用改进的BP神经网络标定对比度的方法.首先,设计了用于对比度标定的BP神经网络模型.然后,利用LM( Levenberg-Marquardt)算法结合缩放法改进神经网络以提高其收敛速度及泛化能力.最后,通过可调对比度目标源装置实验平台,由测量的辐照度得出了对应的图像对比度数据,使该装置可以通过调节辐照度实时获得规定的对比度.与传统BP神经网络方法相比,改进后的BP神经网络收敛速度快,泛化能力强.标定精度比经典BP算法提高了100倍,比最速下降法提高了10倍.训练次数仅需2 876次时,对比度的标定值与目标值的误差最大值是0.01％,训练均方误差收敛为0.000 459 441,测试误差收敛为0.000 467 003,满足了对检验装置中对比度标定的需要.
展开▼
