Research On Steel Scrap Grading Recognition Based On Convolutional Neural Network And Image Information Entropy

With the increasing ratio of scrap steel in the steel industry,it is particularly important to improve the efficiency and accuracy of scrap steel recognition.In this thesis,the scrap steel identification grade classification problem is studied using a combination of convolutional neural network and image information entropy.Aiming at the traditional method of grading quality inspection staff of scrap steel enterprises relying on vision for grading,which has problems related to high work intensity and visual fatigue,this paper explores the feasibility of combining deep learning algorithms and scrap steel grading recognition from the practical application feasibility using current solutions in the field of computer vision.It also proposes an overall scrap steel determination method based on image information entropy,which in turn enables quantitative comparison of field stacked scrap steel and provides an auxiliary method for manual scrap steel grading on site.The specific contents of the paper and the main conclusions include the following.(1)Preliminary pre-processing of the scrap steel images,using different preprocessing algorithms,to make a usable scrap steel image dataset.The preprocessing performed includes: grayscale processing,grayscale transformation,image smoothing,and image denoising of the images.The dataset is also expanded by extracting frames from the video taken at the site to increase the number of datasets by dividing the dataset into different styles for reasons such as insufficient training datasets and difficulty in collecting the data from the scrap pile.Pre-processing is a preliminary processing of image data and a cleaning of the data to improve the quality and quantity of the dataset.(2)After the initial processing of the dataset,the generic scrap images that have been pre-processed are detected using the edge detection operator.It is found that the Canny operator has a good effect on edge recognition of generic scrap steel plates and scrap wheels.The edge sharpening process is also applied to the blurred edges of the large class of scrap steel images obtained by video frame extraction to reduce the impact of extra noise in the images on the calculation of image information entropy.After that,the dataset is labeled using the image annotation tool PASCAL VOC to make it a training dataset and a test dataset.(3)Training is performed in a network built using Tensor Flow and Keras based on Windows 10 OS,and the Faster R-CNN loss function parameters are optimized for effective training.The training results find that the network converges at a manageable speed with good results when the Batch Size is 20,and the selection of candidate regions also achieves better results.The convergence speed decreases rapidly when the Batch Size is larger than 20,which is not conducive to the convergence of the model.Based on the improvement of the smaller scrap steel recognition problem in the improved Faster R-CNN network,the feasibility of solving the reduction of duplicate regions is proposed.(4)An image information entropy calculation method is proposed to determine the grade of scrap steel in the pile site of a scrap steel recycling company.The actual information entropy of the scrap steel pile at the site of the recycling company is obtained by performing actual calculations on the images of the scrap steel pile at the site obtained after pre-processing,and comparing the calculation results of the same scrap steel pile under different color information.Meanwhile,Python code is used to realize two-dimensional grayscale information entropy at the same time as threedimensional RGB image information entropy.And for the problem of computing time,the C++ code is further optimized and analyzed and compared with the Python code,and it is found that the computation time complexity is significantly reduced.Image information entropy can be used in the field of scrap grading to measure the scrap grade at the scrap pile site,and to quantify the image information entropy of standard highquality scrap to obtain specific values as a metric to assist manual grading and improve the accuracy of scrap grading.This thesis provides two scrap grade determination methods by using deep learning related convolutional neural network for scrap steel identification and introducing image information entropy as a scrap determination criterion to assist manual grading,which can provide further reference for scrap steel identification and grading research.