为了使预测器在特定应用环境中的有限字长效应满足应用系统的性能要求，以小目标检测为应用背景，提出了理论和实验相结合的确定TDNLMS(二维归一化最小均方误差)自适应预测器运算字长的方法。同时分析了步长参数、输人数据字长、图像统计特性、预测器支撑区域等因素与TDNLMS自适应预测器权值和迭代运算中间结果量化误差之间的联系，并通过实验对分析结果进行了验证。仿真结果表明，用该方法设计的有限精度预测器，其小目标检测性能与无限精度预测器十分接近。

The TDNLMS (two dimensional normalized least mean squared error) adaptive filter can be used as a pre whitening filter in small object detection application in digital image. The performance of the digital implementation of the adaptive filter is influenced by the finite precision effect. In this paper, the finite-precision effect of TDNLMS adaptive predictor used for small objects detection is analyzed, and a method of determining computation word length of the digital adaptive filter by experiments is presented. The relation between the word length of the adaptive predictor and the features of the environment, including the step size parameter, input data word length, statistic characteristic of image processed and the support region of the predictor, is discussed. Simulation results are consistent with the analysis. Through the method presented in this paper, a finite precision TDNLMS adaptive predictor is designed, and compared with an infinite precision predictor witch is implemented by double precision floating point numbers. Simulation results shown that the MSE (mean squared error) and the mean SNR (signal to noise ratio) gain produced by the finite precision TDNLMS adaptive predictor is very close to the infinite precision TDNLMS adaptive predictor.