基于自注意力机制的桥梁螺栓检测算法

基于构建的真实桥梁螺栓场景数据集，提出基于自注意力机制与中心点回归（SACPR）的螺栓检测算法. 构建基于真实场景的高质量桥梁螺栓场景数据集，并针对数据不均衡、多样性不够的问题，使用数据增强方法进行数据扩充，从而获得更高的分类精度. 采用基于深度学习框架的SACPR算法检测不同场景下的螺栓，并进行标示. 对螺栓检测准确率进行验证实验，验证SACP算法的有效性. 将试验结果与YOLOv3、Faster-RCNN、RetinaNet这3种算法结果进行对比，发现3种检测方法的识别精度分别为80.56 %、87.71%、93.89%，而所提出的SACPR 算法的识别精度为93.91%，明显优于YOLOv3算法和Faster-RCNN算法；虽然SACPR算法与RetinaNet算法的识别精度较接近，但前者的检测速度是后者的5.6倍.

Self-attention mechanism based bridge bolt detection algorithm

A bolt detection model algorithm based on self attention mechanism and center point regression (SACPR) was proposed based on the real bridge bolt scene data set. Firstly, a high-quality bridge bolt scene data set based on the real scene was constructed, and for the problems of data imbalance and insufficient diversity, data enhancement method was used to expand the data, so as to obtain higher classification accuracy. Secondly, SACPR model algorithm based on deep learning framework was used to detect bolts in different scenes, and label them. Finally, the validity of the proposed method was verified by the verification experiment of bolt detection accuracy. Comparison was conducted with the results of YOLOv3, Faster-RCNN and RetinaNet, and results showed that the recognition accuracy of the three detection methods was 80.56%, 87.71% and 93.89% respectively, while the recognition accuracy of SACPR model method was 93.91%. The accuracy of SACPR model method was obviously better than that of YOLOv3 model algorithm and Faster-RCNN model algorithm. Although the recognition accuracy was almost the same as that of RetinaNet model algorithm, the detection speed of SACPR model method was 5.6 times of that of RetinaNet model.