基于改进的Faster RCNN的仪表自动识别方法

Automatic instrument identification based on improved Faster RCNN

在环境复杂的工业场景中,仪表盘存在类别多、相似性高等问题,导致检测的识别效果较差、准确率不高。针对这一问题,提出了一种基于改进的更快速的区域卷积神经网络(Faster RCNN)的仪表自动识别方法。首先,采用残差网络(Resnet)101代替视觉几何群网络(VGG)16,进行了网络结构简化;然后,引入了特征金字塔网络(FPN),并将其改进为递归特征金字塔网络后进行了迭代融合,输出了特征图;接着,引入了注意力机制模块,根据特征的重要程度,完成了对输出通道权值的重新分配,增强了Faster RCNN对目标的运算能力;提出了改进非极大值抑制算法(Softer-NMS),通过降低置信度来确定准确的目标候选框;最后,采用Mosaic数据增强技术对可视对象类(VOC)2007数据集进行了扩充,对改进后的Faster RCNN模型进行了仪表自动识别的实验。研究结果表明:在相同工业环境下,与传统的Faster RCNN算法模型相比,改进后的Faster RCNN模型准确率为93.5%,较原模型提高了3.8%,mAP值为92.6%,较原模型提高了3.7%,可见该方法在实际生产中具有较强的鲁棒性与泛化能力,可满足工业上对智能检测的要求。

Aiming at the problem of poor detection and recognition effect and low accuracy due to complex detection environment,high similarity of instrument panels,multiple category classification and other interference in industrial scenes,an automatic instrument recognition method based on improved faster regional convolutional neural network(Faster RCNN)was proposed.Firstly,the residual network(Resnet)101 was used to replace the visual geometry group network(VGG)16 for network structure simplification.Then,the feature pyramid network(FPN)was introduced and further improved into a recursive feature pyramid network for iterative fusion,and the feature map was output.Then,the attention mechanism was introduced to realize the weight of the output channel and reassign it according to the degree of importance to enhance the model s computation of the target.A softer non-maximum suppression(softer-NMS)algorithm was introduced to determine the accurate target candidate box by reducing the confidence degree.Finally,Mosaic data enhancement technology was used to expand the visual object classes(VOC)2007 data set,and the improved Faster-RCNN model was used to carry out the instrument automatic recognition experiment.The results show that,in the same industrial environment,comparing with the traditional Faster RCNN algorithm model,the accuracy rate of the improved Faster-RCNN model is 93.5%,which is 3.8%higher than that of the original model,and the mAP value is 92.6%,which is 3.7%higher than that of the original model.It can be seen that this method has good robustness and generalization ability in actual production,and can meet the requirements of intelligent detection in the industry.