Abstract: The hyperbolic echo feature is one of the most important elements for root recognition in ground penetrating radar (GPR) image. By intensive analysis of the hyperbola detected accurately from tree root image by GPR, an improved method was presented to detect the target curve. The new method includes two aspects of content: 1) According to the particularity of GPR image's wavelet spectrum information, an extraction method of region of interest ( ROI) was proposed based on GPR gradient magnitude. Furthermore, in order to make the hyperbolic characteristics in the gradient magnitude diagram more outstanding, the way of differential value was optimized by gradient magnitude method to achieve the rapid extraction of hyperbolic. 2) Hough transform was conducted to detect the hyperbola in ROI. By using location information and amplitude information comprehensively voting, the accuracy of original Hough transform was improved, and the background clutters and false targets can be removed effectively. Noise immunity of the method proposed was strong. In addition, the stability of the method was high. Thereby the accurate extraction of the target hyperbola is realized.