Cracks and bugholes (surface air voids) are common factors that affect the quality of concrete surfaces, so it is necessary to detect them on concrete surfaces. To improve the accuracy and efficiency of the detection, this research implements a novel deep learning technique based on DeepLabv3+ to detect cracks and bugholes on concrete surfaces. Firstly, in the decoder, the $3\times 3$ convolution of the feature fusion part is improved to a 3-layer depth separable convolution to reduce the information loss during up sampling. Secondly, the original expansion rate combination is changed from 1, 6, 12, 18 to 1, 2, 4, 8 to improve the segmentation effect of the model on the image. Thirdly, a weight value is added to each channel of the Atrous Spatial Pyramid Polling (ASSP) module, and the feature maps that contribute significantly to the target prediction are learned and screened. To use this method, a database is built containing 16, $662\,\,256\times256$ pixel images of bugholes and cracks on concrete surfaces. The two defects included in those images are labeled manually. The DeepLabv3+ architecture is then modified, trained, validated and tested using this database. A strategy of model-based transfer learning is applied to optimize and accelerate the learning efficiency of the model. The weights and biases of the Xception part of the model are initialized by the pretrained backbones. The results are 97.63% (crack), 93.53% (bughole) Average Precision (AP), 95.58% Mean Average Precision (MAP) and 81.87% Mean Intersection over Union (MIoU). A comparative study is conducted to verify the performance of the proposed method, and the results demonstrate that the proposed approach performs significantly better in crack and bughole detection on concrete surfaces.

中文翻译：

---

基于深度学习的混凝土表面图像自主裂纹和漏洞检测

裂缝和漏洞（表面气孔）是影响混凝土表面质量的常见因素，因此有必要在混凝土表面进行检测。为了提高检测的准确性和效率，本研究实施了一种基于 DeepLabv3+ 的新型深度学习技术来检测混凝土表面的裂缝和漏洞。首先，在解码器中， $3\times 3$ 特征融合部分的卷积改进为3层深度可分离的卷积，以减少上采样过程中的信息丢失。其次，将原来的扩展率组合由1、6、12、18改为1、2、4、8，以提高模型对图像的分割效果。第三，对 Atrous Spatial Pyramid Polling (ASSP) 模块的每个通道添加一个权重值，学习和筛选对目标预测有显着贡献的特征图。为了使用这种方法，构建了一个包含 16 个的数据库， $662\,\,256\times256$ 混凝土表面上的漏洞和裂缝的像素图像。这些图像中包含的两个缺陷是手动标记的。然后使用该数据库修改、训练、验证和测试 DeepLabv3+ 架构。应用基于模型的迁移学习策略来优化和加速模型的学习效率。模型 Xception 部分的权重和偏差由预训练的主干初始化。结果是 97.63%（裂缝）、93.53%（漏洞）平均精度 (AP)、95.58% 平均平均精度 (MAP) 和 81.87% 平均交集 (MIoU)。进行了比较研究以验证所提出方法的性能，结果表明所提出的方法在混凝土表面的裂缝和漏洞检测方面表现明显更好。