Abstract: Cracks are one of the most important diseases of bridges. Timely and efficient detection and evaluation of cracks is very important to maintain the health of bridges. Aiming at the high integration cost and low detection accuracy of fracture annotation data, an improved DeeplabV3+ model based on attention mechanism and transfer learning is proposed. The model adds attention mechanism to obtain rich context information, improve the learning ability of crack feature channel and reduce the influence of background noise and other features. Then, through the combination of public data set and small sample data set, the source domain data set and target domain data set are established for migration learning, so as to reduce the impact of too few training samples on detection performance. The experimental results show that the improved DeeplabV3+ model has achieved good detection effect on bridge crack detection, and the detection accuracy has reached 93.3%, which is 3 percentage points higher than the original model. The transfer learning training model achieves high detection accuracy on small sample data, which can save a lot of labeling costs.

Key words: crack detection, small sample learning, transfer learning, semantic segmentation, attention mechanism

摘要： 裂缝作为桥梁最主要的病害之一，及时高效地发现和评估裂缝对保持桥梁的健康状况至关重要。针对裂缝标注数据集成本高、检测精度低等问题，提出了一种基于注意力机制和迁移学习的改进DeeplabV3+模型。该模型通过添加注意力机制来获取丰富上下文信息，提高裂缝特征通道的学习能力，降低背景噪声影响；通过公共数据集和小样本数据集组合建立源域数据集、目标域数据集以供迁移学习使用，以此来降低训练样本过少对检测性能的影响。实验结果表明，改进DeeplabV3+模型对桥梁裂缝检测获得了较好的检测效果，检测精度达到了93.3%，较原始模型提高了3个百分点；通过迁移学习训练模型在小样本数据上取得较高的检测精度，可节省大量标注成本。

关键词: 裂缝检测, 小样本学习, 迁移学习, 语义分割, 注意力机制