A Lightweight Detection and Recognition Framework for cigarette laser code

Abstract

Automatic detection and recognition of laser codes on cigarette case is important in distinguishing the authenticity of cigarettes. However, detecting and recognizing cigarette laser codes presents a challenging industrial problem due to the intricate background of cigarette images. In this paper, a Lightweight Detection and Recognition Framework (LDRF) is proposed to detect and recognize cigarette laser code. The LDRF model consists of the Lightweight Detection Network (LDNet) and the Lightweight Recognition Network (LRNet). In the LDNet stage, a lightweight feature extraction network is proposed to extract features of the cigarette code area. Furthermore, a bidirectional feature pyramid network (BiFPN) feature fusion structure is introduced to tackle the multi-scale feature fusion challenge in cigarette code detection scenarios. Notably, alignment and normalization of all features channels are conducted to reduce computational requirements in the post-processing stage, enabling precise detection performance while significantly reducing parameters and computational complexity. In the LRNet stage, an integrated network architecture is designed to enhance the fusion of visual and temporal features. Furthermore, a combination of bidirectional temporal convolutional network (BiTCN) and Transformer is employed in the feature extraction stage to differentiate between background and characters, as well as capture the interdependence among different characters. Specifically, DownSampling is utilized to adjust the size of input images and Merging or Combining methods are applied at each stage to capture multi-level features. Experimental results demonstrate that the proposed LDRF method provides better performance than state-of-the-art models, and achieves trade-off between accuracy and speed.