The necessities of pattern recognition systems are much increased in many applications. According to their applications, they may be divided into two major groups; speech \& speaker recognition and image-based recognition. Image-based pattern recognition systems, we called image understanding systems, are widely used. In these systems, edge extraction is indispensible in order to classify the given input images. There has been a number of edge extraction techniques. Most of existing edge extraction techniques have some problems, especially for noisy images. We propose a new technique which works well for noisy images. It is based on statistical testings and edge searching procedure. It has additional capability of detecting lines as well as step edges, and can be applicable to various types of Images. Its algorithm is rather complex compared with several existing techniques such as method of using simple edge operators; however, these techniques requires some preprocessing \& postprocessing to have good results. On a while, this proposed technique won``t use any preprocessing and does not need postprocessing, for example edge thinning, etc., which makes computing time to be much saved. Also, most of existing edge extraction techniques have difficulties in selecting a threshold value, whereas for this proposed technique its selection is very easy and it has been found experimentally that the resulting edges are rather insensitive to the particular choice of threshold value, regardless of types of images and amount of noise.