一维条形码是一种用于商品识别及跟踪的数据编码信息载体，由一组特定宽度的黑白矩形条遵循预先设定的编码规则排列组合构成。随着经济社会的飞速发展，条形码已经在商品零售、物流快递、医疗卫生等诸多和大众生活息息相关的领域有着广泛应用，有重要研究价值。条码检测定位是识别条码信息的必要前提，同时，实现自动且精确的条码定位，也有利于提升条码识别的速度和准确率。在当前的条码检测方法中，传统的基于图像几何特性的方法对应用场景有较高限制，当条码图片分辨率高或者背景纹理复杂时背景噪声会产生大量额外计算。而现有的基于深度学习的条码检测方法，将条码的检测定位和角度检测分阶段实现，这一方式存在以下不足：首先，模型分阶段完成条码的初步定位和角度检测，模型结构复杂，训练和推理的时间成本高。其次，条码的检测定位和角度检测过程都会存在误差，分阶段实现会导致误差累积，导致模型性能变差，不利于后续的解码识别。为了解决上述问题，本文提出一种单阶段的准确条码检测模型，在完成条码定位的同时计算获得条码倾斜角度，提高检测速度，减少计算误差，提高模型鲁棒性。本文的主要工作和贡献如下：1. 提出了一种单阶段的一维条码检测定位算法，本方法在检测算法中设计了角度检测模块，在完成条码定位同时，获得条码角度信息实现准确条码检测。解决了前人方法中分阶段实现检测定位和角度矫正速度慢的问题，有效提升了条码检测速度。2. 设计了颈部特征增强模块, 该模块通过对骨干网络的多个特征层分别使用RFB 提取特征并进行特征融合实现。此外，在骨干网络中引入混合注意力机制 CBAM 以提高网络对特定通道和特征区域的关注度。该设计有助于提升模型对小尺度条码的检测精度和角度检测的准确性。3. 为验证模型的有效性，本文在公开数据集 ArteLab 和 Muenster 以及自建的条码数据集上设计了充分的对比实验。本研究的实验结果显示，与当前最好方法相比本文方法在检出率相近的前提下，检测速度显著提升，并且有更高的条码检测精度，证明了本文方法的优越性。
The 1D barcode is a data encoding information carrier used for product identification and tracking. It consists of a set of black and white rectangular bars of a specific width which is arranged and combined according to the encoding rules. With the rapid development of the economy and society, barcodes have been widely used in many fields which is closely related to public life, such as commodity retail, logistics express, healthcare, and so on, thus having important practical research value.Barcode detection and positioning is a necessary prerequisite for identifying barcode information. At the same time, automatic and accurate barcode positioning is conducive to improving the speed and accuracy of barcode recognition. Therefore it’s evident that accurate barcode detection is an essential and significant preparatory step for barcode identification. Existing barcode detection methods can be divided into two categories. The traditional method based on the geometric characteristics of the image has high restrictions on the application scene. When the resolution of the barcode image is high or the background texture is complex, the background noise will lead to massive unnecessary calculations. However, the existing barcode detection method based on deep learning implements detection and angle prediction of the barcode in stages, which has two obvious shortcomings. First, the model completes the preliminary positioning and angle detection of the barcode in stages, and the model structure is complex. So the time cost of training and reasoning is high. Secondly, there will be minor deviations in the barcode detection and angle prediction, and implementation in stages leads to deviation accumulation, resulting in poor model performance, which is not conducive to subsequent decoding and recognition.To solve these problems, we propose a single-stage accurate 1D barcode detection model in this thesis, which calculates the barcode angle and completes the barcode positioning at the same time, resulting in higher detection speed and more accyrate barcode detection. The main work and contributions are as follows:1. We propose a single-stage 1D barcode detection and positioning algorithm. Specifically, we design an angle detection module to synchronously complete the barcode positioning and angle prediction, effectively improving the barcode detection speed.2. We introduce the neck feature enhancement module, which is implemented by using RFB to extract features and perform feature fusion on multiple feature layers of the backbone network. In addition, the hybrid attention mechanism CBAM is introduced in the backbone network to extract valuable features more efficiently. This design helps improve the model detection accuracy for small-scale barcodes and angle detection.3. To verify the validity of the model, we design sufficient comparative experiments on the public datasets ArteLab and Muenster and the self-built barcode dataset. The experimental results show that compared with the current best method, the detection speed of our model is significantly improved under the premise of a similar detection rate, and the barcode detection accuracy is higher, which proves the superiority of our proposed single-stage 1D barcode detection approach.
