In order to effectively improve the accuracy of crack identification and crack characterization method, this research has focused on the problem of dark fields, low contrast and high noise in the computed tomography (CT) images of HMX-based polymer bonded explosives (PBXs). We propose a new algorithm. By using the density and local-orientation of cracks, the algorithm first preprocesses the source image by integrating the non-local mean algorithm, the Laplacian sharpening algorithm and the Gamma sharpening algorithm. Then, it adopts the statistical method to extract directly cracks’ feature-points, which is used to derive the positions, orientations and thicknesses of cracks via the local orientation among these points measured by the Mahalanobis distance. Finally, it copies the gray values of pixels from the source image, to realize fine recognitions of continuous cracks from CT images of HMX-based PBXs. Comparative experiments were conducted, on CT images of thermal-force compound loading cracks from three PBXs, by this proposed Mahalanobis distance based crack extraction(MDCE) algorithm, Canny edge detection algorithm and phase consistency method, respectively. Results show that this proposed MDCE algorithm can extract various cracks clearly and accurately cracks from CT images with low quality, which indicates the high effectiveness of this method and enhances the ability of crack recognition and characterization.