基于深度学习算法的水尺刻度提取技术

Algorithm of Draft Scale Detection Based on Deep Learning

针对基于传统图像处理的水尺刻度检测算法受天气、光照等诸多环境因素影响较大,提出一种语义分割和目标检测结合的检测方法。首先在语义分割模型SegNet中嵌入注意力机制提高水域图像分割精度,然后将水尺区域图像送入改进的YOLOv3网络进行特殊字符的识别,并结合水岸线位置推算不完整字符长度,得到完整水尺刻度。实验结果表明:该水尺检测算法能准确定位水尺且水尺刻度提取准确度达到95%以上。

In view of the influence of weather,illumination and other environmental factors on the traditional water gauge scale detection algorithm based on traditional image processing,a detection method combining semantic segmentation and target detection is proposed.Firstly,attention mechanism is embedded into the semantic segmentation model SegNet to improve the segmentation accuracy of water area image.Then,the image of water gauge area is sent to the improved YOLOv3 network for special character recognition,and the incomplete character length is calculated by combining with the position of water shoreline to obtain the complete scale of water gauge.The experimental results show that the algorithm can locate the draft accurately,and the accuracy of draft scale extraction is more than 95%.