Research On Industrial Character Detection Method Based On Convolutional Neural Network

At present,the image has become one of the information obtained from the industrial scene.For many elements such as intelligent production process and intelligent production control in industrial intelligence,image information promotes the innovation of these elements,greatly improves the efficiency of industrial production and reduces the cumbersome degree of operation of staff.The character information in the image is closely related to the production,inquiry,tracking and checking in the industrial production process.Therefore,extracting effective character information from industrial scene images is of great significance to industrial intelligence.Based on the convolutional neural network,this paper studies the problems in the process of industrial character detection,such as circuit board character dataset and steel pipe serial number dataset.The accuracy and speed of industrial character detection are effectively improved.The main contents of this paper are as follows:Firstly,traditional character detection algorithm,based on machine learning and deep learning,are analyzed.Equipment selection and dataset production for character collection are completed in the two scenarios of circuit board character and steel pipe serial number.The experimental results show that the overall recognition rate of circuit board character under the traditional algorithm is 39.09%.Traditional algorithms are less effective in industrial scenes with complex environments.Character detection algorithms,based on deep learning,have better detection performance.But they also have the disadvantages of lower detection accuracy and slow detection speed.Secondly,because of the missed detection and false detection in Faster R-CNN,an improved character detection method,called by Faster R-CNN-FF,is proposed.In the Region Proposal Network(RPN),the selection of anchor frames is optimized by the aspect ratio characteristics of character targets in the dataset.Then,Multi-resolution feature fusion is introduced in ROI Pooling layer to extract more convolutional features of candidate regions,and save more semantic information and location information.The experimental results show that the overall recognition rate of the circuit board character under the above algorithm is89.09%.The overall recognition rate of the steel pipe serial number is 77%.The above algorithm is obviously better than the traditional algorithm and single-stage deep learning algorithm.It has achieved better detection results on the circuit board character dataset and steel pipe serial number dataset.Thirdly,missing detection and slow detection speed often exist in the Fast-R-CNN-FF model.The information between character sequences can not be made full use of.A character detection algorithm,based on PSENet,together with the improved multi-head self-attention CRNN,is proposed.Character detection process separates the character area positioning process from the recognition process.Shufflenetv2-FPN is used to replace Res Net50-FPN as the feature extraction network of PSENet to realize character location.The MHSA module,embedded with position information,is embedded in Res Net18.Res Net18 is used as the feature extraction network of CRNN.And the auxiliary loss function Center Loss and CTC Loss are combined for model training to complete character recognition.The experimental results show that the overall recognition rate of the circuit board character reaches 97.27%.The overall recognition rate of the steel pipe serial number reaches 95%.And the detection process is about 70 ms.Based on the above algorithm,the character detection requirements in the industrial environment can be well met.Fourthly,the character positioning and character recognition algorithms are connected in series and then packaged to build an industrial character detection platform.The platform has been practically and reliably applied.