Abstract:

Crack detection is significant important for the safe operation of facilities and equipment. However, some problems exist in current crack detection methods：the data annotation process is cumbersome; the training process is time-consuming; and the detection speed is slow. In order to solve these problems, a weakly supervised crack detection algorithm is proposed. The algorithm uses the features extracted by the convolutional neural network to generate a saliency map containing the position information of the crack in the image, thus achieving crack detection. The experimental results show that the crack detection algorithm based on improved saliency map solves the difficulty in data annotation, and has a pixel-level detection effect while only need to learn image categories. The training time is short, only needs 23 minutes using two k80 GPUs, at the same time, achieving 52% average precision on the homemade data set and has a detection speed of 105 frames per second, which verifies the effectiveness of the algorithm.

Key words: crack detection, convolutional neural network, saliency map, pixel-level detection, real-time detection

摘要：

裂纹检测对于确保设施和设备的安全运行非常重要。然而现有的裂纹检测方法数据标注困难，且训练时间长，检测速度慢。针对这些问题，提出一种基于改进显著图的裂纹检测算法，应用卷积神经网络提取的特征计算出图像中裂纹的位置，从而实现裂纹检测。该算法解决了裂纹检测算法数据标注困难的问题；训练时间短，使用两块k80显卡训练时间仅为23 min；检测速度快，具有105 帧/s的检测速度；同时在自制数据集上达到了52%的平均精度，验证了该算法的有效性。

关键词: 裂纹检测, 卷积神经网络, 显著图, 像素级检测, 实时检测