改进YOLOv3的桥梁表观病害检测识别

针对基于目标检测方法的桥梁表观病害检测存在检测精度低、误检率和漏检率高的问题，提出一种改进YOLOv3的高准确率桥梁表观病害检测识别方法。为实现局部特征和全局特征有效融合，在YOLOv3的检测层中添加固定分块大小的池化模块，并在YOLOv3的特征提取网络中引入了DenseNet密集型连接网络结构以增强桥梁病害特征在网络中的传播和利用效率，提高检测效率，采用数据增强技术来扩充样本图像以解决现有桥梁病害数据集样本数量不足的问题。实验结果表明，改进后的YOLOv3在桥梁表观病害检测上的平均准确率比原YOLOv3提高了3.0%,且模型训练时间减少了33.2%,同时降低了对桥梁表观病害检测的误检率和漏检率。

Bridge apparent disease detection based on improved YOLOv3

To solve the problems of low detection accuracy, high false detection rate, and high missed detection rate in bridge apparent disease detection based on target detection method, a recognition method with high accuracy of bridge apparent disease detection based on improved YOLOv3 is proposed. A pooling module with a fixed block size is added to the detection layer of YOLOv3 to realize effective fusion of local features and global features. To enhance the transmission and utilization efficiency of bridge disease features in the network and improve the detection efficiency, a DenseNet dense connection network structure is introduced in the feature extraction network of YOLOv3. To deal with the insufficient number of samples in the existing bridge disease data set, data enhancement technology is used to expand the sample images. The experimental results show that the mean accuracy precision (mAP) of the improved YOLOv3 on bridge apparent disease detection is increased by 3.0% and the model training time decreased by 33.2%, with a reduced false detection rate and a lower missed detection rate.