Abstract: Due to the issues of blurred background,small character areas,inconsistent engraving positions and paint occlusion in the images of cast pipe characters collected from industrial sites,the existing models struggle to meet the detection accuracy requirements of industrial environments.To address these issues,an improved Cascade Region Convolutional Neural Network (RCNN)for cast pipe character detection was proposed.The improvements were made to the feature pyramid by introducing the Small Target Enhancement Feature Pyramid Networks (STE-FPN),which utilized the feature enhancement capability of multi-scale feature fusion to enrich the feature information of small target characters on cast pipes.The Self-Calibrated/Pooling ResNeSt (SCP-ResNeSt) was introduced as the feature extraction network,utilizing self-calibrated convolutions and pooling operations to enhance the efficiency of extracting complex background features of cast pipe characters.The improvements were made to the cascade structure by introducing a Mask branch structure,which adaptively detected character areas and removes interference regions,thus optimizing the detection results.The improved algorithm was tested on the cast pipe dataset,and mean Average Precision (mAP) was 99.1%,which increased the precision by 2.3% compared to the Cascade RCNN.The precision showed that the improved performance was superior to the original algorithm.

Key words: character detection of cast pipes, blurred background, Cascade region convolutional neural network, ResNeSt

摘要： 由于工业现场采集的铸管字符图像存在背景模糊、字符区域占比小、刻字位置不固定、油漆遮挡等问题,导致现有模型的检测精度难以满足工业现场的需求。针对上述问题,提出改进的Cascade RCNN铸管字符检测算法。首先对特征金字塔进行改进,提出融合小目标增强的特征金字塔(STE-FPN),利用多尺度特征融合的特征增强能力丰富铸管小目标字符的特征信息。其次引入自矫正/池化的ResNeSt(SCP-ResNeSt)作为特征提取网络,利用自矫正卷积和池化操作以提升背景复杂的铸管字符特征提取效率。最后对级联结构进行改进,引进Mask分支结构,可以自适应地检测字符区域并去除干扰区域,优化了检测结果。将改进后的算法在铸管数据集上进行测试,其平均检测精度mAP为99.1%,比原Cascade RCNN算法提高了2.3%,得到的精度表明改进后的性能优于原算法。

关键词: 铸管字符检测, 背景模糊, Cascade RCNN, ResNeSt