Automatic threshold generation technique

Abstract

An estimate of a threshold value for feature extraction from image data containing noise and quantized into a plurality of quantization levels is made by filtering, such as recursively averaging n adjacent quantization levels and counting the extrema of the result. The recursive averaging is repeated with increasing values of n until only three extrema are present in the filtered data. The estimate of the threshold value is refined by comparing averages or summations of n quantization levels above and below the estimate and altering the estimated threshold value based on a comparison of the averages or summations. By basing both the estimation and refinement processes on the original data and using a value of n which is smaller in the refinement process than in the estimation process, a highly effective threshold value for the feature extraction process is obtained. Adaptive adjustment of n during the estimation process minimizes loss of data as well as computing time.