Aiming at the robustness of traffic sign recognition (TSR) under the interference of occlusion noise, this paper proposes a collaborative recognition method that fuses the segmented cluster sparse coding of integrated features with a twin network. First, the integrated coding is constructed by multi-scale feature fusion (color covariance matrix, shape context, etc.); second, the segmented cluster sparse optimization algorithm is designed to enhance the internal structure consistency of the features and suppress the redundant noise; and finally, the lightweight twin network is constructed to achieve efficient feature matching through the sharing of the weight structure and the comparison learning mechanism. Experiments show that: on the TT100K dataset, this method reduces the amount of parameters by 59.6%, improves the FPS by 51.4%, and the original scene recognition accuracy is comparable to that of SOTA; on the occlusion test set, the accuracy and other core indexes are significantly better than that of the optimal baseline model, which verifies its strong fault-tolerance for partial feature missing.