This report proposes a new method which can detect internal weld defect areas appearing in X-ray radiographic films, and verifies its efficiency using actual defect images. The method has robustness against noise and blurred defect images, because each pixel of image votes for an area decided by the direction of gradient vector of the pixel. And the method can classify the image into defective areas and indefective areas by thresholding the voted image. The results using actual images are as follows, (1) more than 90% defects in 125 images are detected, and (2) more than 85% of their characteristic defect shape are found.