The Hough transform (HT), that is a typical algorithm for detecting lines in images, needs considerable computational costs and easily detects pseudo-lines on the real world images, because of the large amount of features generated by their complex background or noise. This paper proposes an improved HT that add a preprocessing to estimate the validity of features to the conventional HT. The feature estimation can remove a lot of inessential features for the line detection using a pattern of $3{ imes}3$ block features. Experiments using various images show that the proposed algorithm saves computational costs by removing 14%~58% of features depending on images and besides it is superior to the conventional HT in valid line detection.