基于改进YOLOv4网络的水表读数识别方法

在远程水表读数自动识别系统中，为减少网络模型参数量，改善受雾化、抖动等干扰的水表复杂场景图像读数识别精度及半字识别问题，提出了一种基于改进YOLOv4网络的水表读数识别方法?该方法利用深度可分离卷积与引入压缩与激发(squeeze-and-excitation, SE)注意力机制的MobileNetv2瓶颈结构，分别替代YOLOv4网络原有的标准卷积和主干网络，并利用加权平均非极大值抑制算法改进预测输出头，形成了一种网络模型参数量明显降低但检测精度不下降的改进YOLOv4网络，同时有效改善了对水表读数“半字”识别的漏检和错检问题；最后基于字符边框定位的水表读数提取方法，实现“半字”准确提取问题?实验结果表明，本文方法与多种网络学习方法相比，模型参数量压缩14.4%以上，读数识别的准确率和召回率对普通场景水表图像分别提升了0.04%和0.05%以上，对受雾化、抖动等干扰的复杂场景水表图像分别提升了0.11%和0.37%以上.

Reading recognition of water meter based on improved YOLOv4 network

In the automatic recognition system for remote water meter readings, in order to reduce the number of network model parameters, improve the recognition accuracy and semi-character recognition of water meter images in complex scenes disturbed by atomization, jitter, etc. a reading recognition method based on improved Yolov4 network is presented. An improved YOLOv4 network is formed by using depth-separable convolution and MobileNetv2 bottleneck structure with the squeeze-and-excitation (SE) attention mechanism to replace the standard convolution and the backbone network, and using weighted average non-maximum suppression algorithm to improve the prediction output head in YOLOv4 network. The improved YOLOv4 network can obviously reduce the number of network model parameters without decreasing the detection accuracy. At the same time, it effectively improves the problem of missing and wrong detection in the recognition of half-word of water meter reading. Finally accurate extraction of half-word in water meter images is solved by the method based on the location of character frame. Experiment result and comparisons with several network learning methods show that the proposed method compresses the model parameters by more than 14.4%, improves the accuracy and the recall rate of reading recognition of water meter images in common scene by more than 0.04% and 0.05% respectively, and increases the accuracy and the recall rate of reading recognition of the water meter images in complex scene disturbed by atomization and jitter by more than 0.11% and 0.37% respectively.