A Reading Recognition Model of Pointer Type Oil-level Meter Based on Improved Resnet

Abstract

At present, the reading recognition of pointer type oil level meter mainly depends on manual observation, which is not accurate due to different scenes and perspectives. And traditional methods based on Hough transform are designed according to expert experience, which is inconvenient to be used by the staffs with little relevant work experience. To improve the automation degree of reading recognition, a model based on improved Resnet and improved Bayesian optimization is proposed in this paper. Firstly, convolution kernel decomposition, more shortcut connection and changeable network structure are adopted to improve Resnet18. Secondly, 3 constrains are added to improve Bayesian optimization to speed up the converge process and reduce network size. Finally, use the improve Bayesian optimization to search the suitable hyperparameter of the network, including initial learning rate, momentum of SGD, L2 regularization factor, filters number of first convolution layer, and the number of residual blocks. Example shows that the improved Bayesian optimization can help to converge faster with a small size of network, and improved Resnet performs the best compared with other classical deep learning network.