融合DBNet与LCNet-SVTR的卷烟鉴别检验返样单自动识别方法OA

Manual processing of tobacco identification and inspection return samples in the tobacco in-dustry is time-consuming,labor-intensive,and error-prone,failing to meet the demand for efficient and accurate data processing.To overcome these limitations,this paper proposes a deep learning-based auto-matic recognition method that integrates the differentiable binarization network(DBNet),a lightweight convolu-tional network LCNet,and the sequence vision Transformer(SVTR).The method first preprocesses return sample images through grayscale normalization,contrast enhancement,scale standardization,and data augmentation to sup-press the adverse effects of scanning noise,uneven illumination,and seal interference on recognition performance.Subsequently,in the text detection stage,the DBNet model,which adopts the idea of differentiable binarization,is employed to jointly learn the text probability maps and an adaptive threshold map,thereby achieving the segmenta-tion and localization of dense short text regions.In the text recognition stage,an LCNet-SVTR recognition model is constructed by combining LCNet and SVTR,which captures fine-grained local character detail while modeling global contextual semantics.Finally,the text recognition results of the return samples are output through post-processing decoding.Experimental results demonstrate that the proposed method achieves a recognition accuracy exceeding 89%and a normalized edit distance above 98%,offering an effective solution for automated document processing in the tobacco industry.