In order to optimize and improve object recognition of the existing robots, a new weight calculation method for interior scene images identification is proposed. The undirected weighted graph is obtained by converting the input scene. Based on surface normal direction, comprehensive determination of surface roughness is carried out using the concave and convex degree index instead of the traditional Boolean decision, which greatly enhances the anti-noise performance and avoids the error propagation amplification. The unknown objects are identified in time based on fast image segmentation algorithm. Experiment results show that the robustness and anti-noise ability of the proposed method are both strong, and the proposed method is better than methods based on normal direction only. Compared with other methods based on deep learning and conjecture, the proposed method has better performance, and it’s more applicable to the practical identification.