基于Faster R-CNN与形态法的路面病害识别

为提高基于图像处理的路面表观病害检测识别效率及精度，引入目标检测中的快速区域卷积神经网络（Faster Region Convolutional Neural Network，Faster R-CNN）算法以快速识别病害种类、位置与面积；针对已提取的带边框裂缝病害区域，采用基于VGG16迁移学习与模型微调的CNN与50%重叠率的滑动窗口定位裂缝骨架，进而利用形态法操作提取裂缝形态，计算其长度与宽度；针对Faster R-CNN算法在病害种类识别时漏检率低但误检率偏高的问题，引入精确率、召回率和F₁分数指标对算法进行评估，并根据F₁分数最大值确定相应的病害框像素面积及置信度阈值来降低误检率，以适应路面表观病害多样化的应用场景。运用开发的病害识别算法对广东一高速公路路面进行表观检测。结果表明：所提方法对典型裂缝图片的识别效率及精度均高于单独应用CNN滑动窗口和传统形态法的全局图像处理方法；对分段的裂缝边界框进行合并，且病害框像素面积及置信度阈值取优化值后，横向裂缝精确率由合并前的0.861提升至合并后的0.918，横向及纵向裂缝误检率则分别由调整前的20.4%和23.8%下降至调整后的8.2%和6.9%，漏检率则稍有提高。基于Faster R-CNN、CNN及形态法的路面病害识别方法具有工作高效、漏检率低的优点，在引入评估指标、最优病害框像素面积与置信度阈值后，病害误检率也大幅降低，具有潜在工程应用价值。

Pavement Distress Detection Based on Faster R-CNN and Morphological Operations

To improve the efficiency and accuracy of image-based pavement distress detection as well as quickly identify the type, location, and magnitude of the distress, the Faster R-CNN algorithm for object detection is introduced. A convolutional neural network (CNN) based on VGG16 migration learning and model fine tuning was employed in an extracted crack area with a bounding frame to locate the crack skeleton with a 50% overlap sliding window. Then, morphological operations were conducted to extract the crack skeleton and calculate its length and width. To improve the performance of Faster R-CNN and evaluate the effectiveness of the integrated algorithms where a low misdetection rate but high false-detection rate is likely, the precision, recall, and F₁ score were introduced. The maximum F₁ score was used to determine the pixel area of the distress frame and the corresponding confidence threshold thus reducing the false detection rate and adapting to the diverse scenarios of pavement surface distress. The rapid pavement distress detection algorithm was applied to an expressway in Guangdong, China. The test results on typical crack sample images show that the proposed method is more efficient than full-field image processing methods such as CNN with sliding window and traditional morphology operations. As the segmented crack bounding frames were merged and adjusted, and the optimized pixel area and confidence threshold for the distress boxes were considered. It was observed that the precision rate of the transverse crack increases from 0.861 to 0.918, whereas the false detection rates of the horizontal and vertical cracks decrease significantly from 20.4% and 23.8% to 8.2% and 6.9% before and after adjustment, respectively. The proposed pavement distress detection method integrating Faster R-CNN, CNN, and morphological operations has the advantages of high efficiency and low misdetection rate. Moreover, the false detection rates are greatly reduced by introducing the evaluation method and the thresholds of pixel area and confidence value, which indicates the engineering application potential of the proposed method.