Research On Intelligent Identification Algorithm Of Slab Number In Heating Furnace Based On Deep Learning

Slab number is the only identification of slab in hot rolling production process of iron and steel enterprises.Before the slab is pushed into the heating furnace,the number of each slab is mainly identified by manual method,which is of high labor intensity and prone to errors.Therefore,the automatic identification of slab number based on artificial intelligence technologies with image processing and deep learning is of great significance to reduce labor intensity and improve the informatization and intelligence level of the hot rolling mills in iron and steel enterprises.In this thesis,aiming at the problem of slab number automatic recognition in the heating furnace process of hot rolling mills,the automatic construction of datasets,single-character recognition and overall recognition methods of slab number based on deep learning and image processing were studied respectively.Specific research contents are as follows:(1)Aiming at the problem of huge workload and low efficiency of manual labeling of slab number position,an automatic positioning method of slab number area based on transfer learning was proposed.This method first used the EAST text detection model trained on ICDAR2015 dataset to detect the pictures with slab obtained from video,Then,based on the feature that the length-width ratio of the slab size area is fixed,the identified slab size area was post-processed to improve the positioning accuracy to 89.95%.Finally,based on the captured images of slab number,a sample set of images of slab number was constructed.(2)In order to construct the dataset for the single character recognition and the overall recognition of slab numbers,the pictures of slab number were firstly labeled to complete the construction of the overall recognition dataset.Then,based on the characteristics of slab image,an adaptive global binarization method was proposed to binarize the slab image,and the improved traditional vertical projection algorithm based on clustering was proposed,which could increase the accuracy of slab image segmentation from 85%to 98%.At last the split single character was labeled to complete the construction of the single character dataset.(3)Aiming at the single character recognition of slab number,an improved LeNet-5 network was proposed.The experimental result showed that the recognition accuracy of this method could reach 88.4%.Aiming at the overall recognition of slab number,an overall recognition method based on CRNN was developed,and the accuracy of overall recognition could reach 95.03%.(4)Based on the idea of ensemble learning,the ensemble model of overall recognition based on CRNN and single character segmentation recognition based on LeNet-5 was developed,which could improve the accuracy of recognition from original 95.03%and 88.4%before ensemble to 97.4%.