基于改进YOLOv7的水工混凝土结构表观病害检测

Apparent Diseases Detection of Hydraulic Concrete Structures Based on Improved YOLOv7

针对水工混凝土结构表观病害尺度不均、分辨率低、背景干扰大,现有目标检测算法精度和效率较低的问题,提出了一种改进的YOLOv7检测模型,首先,在主干网络的三个特征输出层后加入CBAM注意力机制,从空间和通道两个维度强化网络对目标特征的关注度;其次,在颈部替换路径聚合网络(PANet)为加权双向特征金字塔网络(BiFPN),进一步融合了浅层的位置信息和深层的语义信息,有效改善了表观病害的检测效果,并替换CIoU为SIoU作为定位损失函数,提高了回归的精度;最后,采用生成对抗网络(GAN)等方式对数据集进行加强,并对检测效果进行可视化处理。试验结果表明,改进后的YOLOv7模型收敛更快,分类精度更高,mmAP值达到89.4%,较YOLOv7提高了3.2%,效果优于其他目标检测算法,实现了病害的实时检测。

The apparent diseases of hydraulic concrete structures are suffered from the uneven scale,low resolution and complex background interference,which brings the lower detection accuracy and efficiency in existing object detection algorithm.An improved YOLOv7 detection model is proposed.Firstly,the CBAM attention mechanisms is added to the three feature output layers of the backbone network to make networks pay more attention to target features from two dimensions of space and channel.Secondly,the path aggregation network(PAN)is replaced to the weighted bidirectional feature pyramid network(BiFPN)in the neck network,which further integrates the shallow position and the deep semantic information.CIoU is replaced by SIoU as the localization loss function,improving the accuracy of regression.Finally,the data is strengthened by means of generative adversarial network(GAN),and the detection effect is visualized.The experimental results show that the improved YOLOv7 model has faster convergence and higher classification accuracy,and the mmAP value reaches 89.4%,which is 3.2%higher than that of YOLOv7.The detect effect is superior to other object detection algorithm,and the real-time detection of diseases is realized.