Incomplete Cigarette Code Recognition via Unified SPA Features and Graph Space Constraints.

Cigarette code is a 32-character string printed on a cigarette package, which can be used by tobacco administrations to determine the legality of distribution. Unfortunately, the recognition task for incomplete cigarette code often suffers from lowered recognition accuracy and the destruction of semantic context due to complex backgrounds and damaged characters. This paper proposes an end-to-end recognition network for incomplete cigarette code to improve recognition accuracy and estimate character landmarks. The proposed network first extracts multi-scale features using feature pyramid networks (FPN), then utilizes a spatial attention (SPA) mechanism to yield unified SPA features and integrates them into instance segmentation. This strengthens spatial representation ability and improves the recognition accuracy. A graph convolutional network (GCN) is introduced to construct graph space constraints and calculate character spatial correlations and accurately estimates missing character landmarks. Finally, we employ the Hungarian algorithm to align recognition characters with estimated landmarks and fill missing characters with '*' to preserve the complete semantic context, and produce the final regularized cigarette code. The experimental results demonstrate that our proposed network reduces time consumption and improves recognition accuracy, surpassing the state-of-the-art methods.