结合特征复用注意力与精细化分层残差的 细微裂纹密集连续检测

细微裂纹的高效识别对结构体早期故障诊断具有重要意义。图像分割等方法在处理复杂且带有断裂的细微裂纹时难以达到满意效果。因此,将细微裂纹的识别问题转变为密集连续的中心点预测问题,利用精细化分层残差模块构造特征提取器并结合具有特征复用的注意力模块提出一种细微裂纹检测方法。首先使用相同的矩形框沿裂纹轨迹密集连续地标注;其次对不同的精细化分层残差模块进行消融实验,得到有利于细微裂纹特征提取的骨干网络;最后结合具有特征复用的注意力模块与骨干网络对比了六种不同的特征复用方式。实验结果表明,本文方法的最高和平均精度分别为61.0%和54.7%,与原模型相比分别提升4.9%和6.3%,成功识别细微裂纹及其局部断裂区域并抑制背景干扰。

Densely continuous detection of micro cracks based on feature reuse attention and refined layered residual

The effective identification of micro cracks is of great significance to the early fault diagnosis of structures. The image segmentation method and other methods are difficult to achieve satisfied results in the detection of micro cracks with complex shapes and broken area. Therefore, transforms the problem of micro cracks identification into a series of dense and continuous central point prediction. A feature extractor is established by using the refined layered residual module, and the feature reuse attention module is also utilized to propose a micro cracks detection method. Firstly, the same rectangular bounding box is used to label the crack track densely and continuously. Secondly, the ablation experiments are implemented on the different refined hierarchical residual module to obtain the backbone network which is conducive to the feature extraction of micro cracks. Finally, six different feature reuse methods are compared by combining the attention module with feature reuse and backbone network. Experimental results show that the highest and average accuracy of the proposed method are 61. 0% and 54. 7% , respectively, which are 4. 9% and 6. 3% higher than the original model. The proposed method successfully identifies the micro cracks and their local broken areas, and suppresses background interference in practical application.