面向多尺度多方向螺栓的检测算法

在工业建设中,螺栓零件是关键的连接件之一,通常用于连接大型机器设备和构件,如钢结构、桥梁、高速公路、建筑、石 油管道等。 其安装状况的优劣将直接关系到整个设备或结构的稳定性和可靠性。 然而,螺栓的安装位置常常处于狭窄、复杂的 环境中,利用人工检测不仅难度大,效率低,而且容易出现误判和漏检的情况。 为此本文以 Faster R-CNN 为基础开展螺栓零件 的识别研究。 针对螺栓零件检测的难点,提出了一种基于多尺度多方向螺栓的检测算法。 首先对采集到的图像进行扩增,以提 高数据集的多样性;其次,通过改变主干网络增强模型对特征信息的敏感程度,再利用多尺度融合模块加强模型对小目标的检 测;在预测框生成阶段,提出自适应旋转区域建议网络,以获取最优预测框;最后,针对多方向检测中出现的边界不连续的问题, 通过 Gaussian Wasserstein 距离和焦点损失作为损失函数来代替传统的 Smooth L1 损失函数。 螺栓零件的识别实验结果表明,改 进后的 Faster R-CNN 模型 mAP 值能达到 87. 4%,相比于原始 Faster R-CNN 模型 mAP 值提升了 7. 6%。 通过消融实验可以得 出,改进后的 ResNet50 网络相较于原始 ResNet50 网络的 AP 值提升了 0. 2%。 与其他旋转检测模型在相同数据集上进行比对 得出,本文提出的模型 AP 值更高,鲁棒性更好。 本文所提出的模型可以解决螺栓零件在识别任务中因拍摄角度和复杂环境出 现的问题,缓解了因图像尺度和旋转边界不连续带来的问题。

Detection algorithm for multi-scale and multi-directional bolts

In industrial construction, bolt components are key connectors commonly used to join large machinery and components such as steel structures, bridges, highways, buildings, and oil pipelines. The quality of their installation directly affects the stability and reliability of the entire equipment or structure. However, the installation of bolts often takes place in narrow and complex environments, making manual inspection difficult, inefficient, and prone to misjudgment and omissions. In this study, we conducted bolt component recognition research based on the Faster R-CNN framework, aiming to address the challenges in bolt detection, we propose a detection algorithm based on multi-scale and multi-directional bolts. Firstly, we augment the collected images to enhance the diversity of the dataset. Secondly, we enhance the sensitivity of the model to feature information by modifying the backbone network, and utilize a multi-scale fusion module to improve the detection of small targets. In the stage of generating bounding boxes, we introduce an adaptive rotation region proposal network to obtain optimal bounding boxes. Finally, we address the issue of discontinuous boundaries in multi-directional detection by employing the Gaussian Wasserstein distance and focal loss as the loss functions instead of the traditional Smooth L1 loss. Experimental results for bolt component recognition demonstrate that the improved Faster R-CNN model achieves a mAP (mean average precision ) value of 87. 4%, which is a 7. 6% improvement over the original Faster R-CNN model. Through ablation experiments, it is observed that the improved ResNet50 network achieves a 0. 2% increase in AP (average precision) compared to the original ResNet50 network. Comparisons with other rotation detection models on the same dataset reveal that the proposed model has higher AP values and better robustness. The model presented in this study effectively addresses the challenges posed by the shooting angles and complex environments in bolt component recognition tasks, mitigating issues caused by image scale and discontinuous rotation boundaries.