Classification And Statistics Of Scrap Steel Based On An Optical Image YOLO Algorithm

Scrap steel is a kind of recyclable resources,and it bears the most recycling value in solid waste.It is also one of the raw materials for steelmaking.In the process of recycling scrap steel,the first problem that the company encountered is the sorting of the scrap steel.But there is no such mechanized equipment system for processing all kinds of scrap steel.This is mainly because the types of scrap steel are extremely diverse,but there is no unified standard for classification.However,it is necessary to distinguish the types of scrap steel for subsequent processing.At present,this kind of work is mainly done manually on the site.This will lead to misjudgment of scrap steel caused by subjective factors,repetitive labor and visual fatigue.In the end,the above-mentioned problems will eventually lead to a large amount of economic loss of the company.This thesis employed optical image processing methods to solve the problem of automatic identification and classification of scrap steel.First of all,a related algorithm is designed to identify and classify scrap steel,and then a designed database system was used to store the statistical results of all types of scrap steel.The realization of this task has certain positive significances for many aspects,such as identifying the types of scrap steel in a certain recycling station and the subsequent treatment of scrap steel.The algorithms used in the thesis include both image processing techniques and the principles of convolutional neural networks.By comparing several mainstream network models,the YOLOv3 network model was selected as the basic model of this thesis for its real-time performance and detection accuracy.This thesis had improved the structure of this network,which used as the backbone network of the scrap classification system.The main tasks of this thesis were as follows:(1)The first step was to preprocess the optical image of the scrap steel,and then constructed the training data sets of the convolutional neural network.The pre-processing employed Gaussian filter to reduce the noise of these scrap images,detected the edge of these images by certain algorithms,extracted the contour of these objects,calculated the circumcircle of the closed contour in these images,and divided it into blocks according to these circles.After processing these images into blocks,there will be a lot of pictures with only one object in them.This method has two merits: one is that it can increase the number of the data sets,and the other is that these pictures can be directly input into the convolutional neural network.Finally,this thesis puts all the photos in the annotation system to annotate the information of the objects inside.(2)The optimized YOLOv3 network is designed according to the characteristics of the optical images of scrap steel.This new network can solve the problem that the original network is not accurate when recognizing the small targets.This thesis adds a feature map of different sizes,which optimizes the network structure of multi-scale feature fusion in YOLOv3.YOLOv3 has three feature maps of different sizes.It merges the last three convolution modules,but this thesis fuses the last four convolution modules.In other words,a new 104 × 104 feature map is added.In this way,the dimension of tensor can be expanded and the detection accuracy of small objects can be improved.(3)This thesis optimizes the loss function of the YOLOv3 network,which can classify the targets in these scrap images.Io U cannot reflect the distance and intersection between the prediction box and the real box,so the value of the GIo U is used to solve this problem;Because the traditional formula of cross entropy function has the same weight for all samples,this thesis uses the focus loss model,which can promote the weight of difficult samples.Finally,this thesis inputs the same training sets and test sets into the improved network model,which is improved based on the characteristics of the scrap optical images.Comparison index includes training effect,training performance and detection effect.After 7000 rounds of training,the stable value of the loss function of the improved model is 8.The stable value of the model before improvement is 11,and the value has decreased by 3 now.The value of AP has increased from 90.2% to 85.05%.The value of m AP has increased from 82.73% to 94.61%.In conclusion,the new YOLOv3 network model provides a research direction for scrap detection.