Abstract A novel node matching algorithm based on contour shape characteristics is introduced for accurately separating touching rice kernels. The original images of touching rice kernels are obtained by a scanner and preprocessed by de-noise, segmentation and contour extraction operations. The extracted contours are then smoothed by convoluting with Gaussian kernel function. Curvature analysis is used to detect characteristic touching points on the boundaries. The first-derivative of the curvature curve is taken to find its local peaks. The computed extremum of curvature correspond to the touching nodes along the original boundaries. Finally, the node matching rules including the confidence radiation region, the shortest distance, the length limitation of splitting line, etc., are proposed to determine an appropriate splitting line between related two of those nodes. The rules are key procedures for dealing with the problems of splitting complex touching kernels, and thus the process of how to determine the splitting line between touching kernels is detailedly discussed. One hundred scanning images with different shapes and sizes of rice kernels are used to estimate the robustness of the algorithm. Experimental results are encouraging that the proposed algorithm is not influenced by the exogenous parameters of rice kernels and can be used to effectively split kernels touching in a very complex way. The proposed methods can eliminate the traditional limitations of the manual placement of rice samples in a non touching manner before image acquisition and implement automatic system for the subsequent inspection of the appearance quality parameters of rice.