Research On Target Detection And Location Method Of Tank Feeding Port Based On Deep Learning

The reduction process of metallic magnesium is an essential step in the smelting process of metallic magnesium.In the reduction process of metal magnesium,the use of object detection technology to identify and accurately locate the feeding port targets of the metal magnesium reduction tank becomes the foundation for achieving automated feeding in the magnesium reduction process.This article introduces the current development status and related basic knowledge of deep learning and object detection in recent years.Based on the actual production scenarios of magnesium reduction processes,a dataset of reduction tank feeding ports was established using on-site real image samples.According to the characteristics of the dataset and the actual needs of the project,the existing algorithm of the actual running target detection and location model is improved and optimized.Based on the semi supervised learning method,without reducing the accuracy,the workload of manual annotation is greatly reduced and the efficiency of annotation is improved.The main research work is as follows:Firstly,optimize the YOLOv3-Tiny object detection algorithm applied on devices to address the issues of insufficient detection accuracy and large errors,in order to achieve higher detection accuracy.The convolution layer is used to replace the maximum pooling layer in the original YOLOv3-Tiny algorithm,improve its original loss function,add Ghost module and CA attention mechanism module in the network structure,and improve the detection accuracy of the target detection algorithm.The experiment shows that the detection accuracy of the improved YOLOv3-Tiny object detection algorithm on the target dataset of the magnesium reduction tank feeding port has been improved from 79.8% to 90.9%,further improving the docking success rate in the automatic feeding process.Secondly,aiming at the problem that a large number of labeled data are required to participate in the training process of supervised target detection,a semi supervised target detection algorithm based on YOLOv5 active self training is proposed combined with the idea of semi supervised learning.For problems such as insufficient detection accuracy and inaccurate positioning of YOLOv5 itself,it was optimized by adding attention mechanism and improving the target loss function.Through experiments,the detection accuracy of YOLOv5’s target detection algorithm was improved from 84.5% to 91.0%,with better detection accuracy and faster detection speed than YOLOv3-Tiny.In view of the quality defects of the pseudo tags generated by the self training algorithm,the quality of the pseudo tags generated is improved and Semantic information in the training data set is enriched by adding a selection strategy of active learning.The experimental results show that,using only 10% of the annotated data from the original feeding port dataset,the improved YOLOv5 object detection algorithm improves the detection accuracy from 88.4% to 91.4%,an increase of 3 percentage points,verifying the effectiveness of the improved method in this paper.