Feature point extraction is the key technology for visual detection and location of weld seam, especially for micro gap in commercial production. Most of the current methods have considerable errors and cannot guarantee the highly-required extracting precision, even fail to recognize the position of micro gaps. Based on autonomous threshold value, the improved median filtering algorithm and feature points extraction algorithm are proposed to deal with the scanning image. Firstly, on the basis of traditional median filtering, the range of threshold value is set by calculating regional mean value and variance, this method is good at protecting the image detail of narrow gap as well as removing noise points. Then, a new method named magnifying details by threshold value is proposed. This method enlarges the gap between feature points of micro gap and nearby data, which enhances seam image details and makes the process of extraction easier. Finally, the error is reduced to 1/ 5　 of that before utilizing time-domain analysis. Experimental results show that this method meets the requirement of high precision, which can greatly recognize the weld seam varying from 0.1 mm to 0.5 mm with an error less than 0.08 mm. It also has the advantages of good adaptability, strong anti-interference ability along with great practical significance in the field of automatic welding of narrow gap.