Abstract: To address the problems of small scale, reflection, multi-angle orientations, and background texture interference in industrial oil-label specification strings, as well as the difficulty of conventional two-stage methods in balancing end-to-end detection-recognition accuracy and system efficiency, an end-to-end oil label text detection and recognition method was proposed. Firstly, a detection-recognition dual-stream decoupling structure was built to reduce feature competition between text localization and character recognition. Secondly, a prediction-driven differentiable rotated cropping module was designed to align rotated candidate regions and generate recognition inputs. Finally, a stable joint training strategy was introduced to improve end-to-end optimization. On the self-built oil label dataset under the Strict-E2E protocol, the end-to-end precision, recall, F1 score, and system speed were 93.75%, 85.49%, 89.43%, and 49.70 FPS, respectively. Compared with Pipeline-SAR, which has a high end-to-end F1 score, the proposed method improves the F1 score by 0.13 percentage points and increases the system speed from 7.44 FPS to 49.70 FPS. Field test results show that the proposed method can support industrial-camera streaming, model verification, and status determination, making it suitable for online reading and production-line verification of oil-label specification strings.

Key words: end-to-end detection and recognition, rotated text detection, industrial text reading, oil label, industrial visual inspection

摘要： 工业机油标签规格短串存在小尺度、反光、多角度分布和背景纹理干扰等问题，传统两阶段方法难以同时兼顾端到端检测识别精度与系统效率。针对上述问题，提出一种端到端的机油标签文本检测识别方法。首先，构建检测识别双流解耦结构，以减弱检测定位与字符识别任务在共享特征上的竞争；其次，设计预测框驱动的可微旋转裁剪模块，实现旋转候选区域的方向对齐与识别输入生成；最后，引入稳态联合训练策略，以提高端到端联合优化的稳定性。实验结果表明，在自建机油标签数据集Strict-E2E评测中，端到端精确率、召回率、F1值和系统速度分别为93.75%、85.49%、89.43%和49.70 FPS；与端到端F1值较高的Pipeline-SAR相比，F1值提高0.13个百分点，系统速度由7.44 FPS提升至49.70 FPS。现场在线检测结果表明，该方法可接入工业相机取流、型号核验和状态判定流程，适用于机油标签规格短串的在线读取与产线核验。

关键词: 端到端检测与识别, 旋转文本检测, 工业文本读取, 机油标签, 工业视觉检测