基于改进的Inception-ResNet-V2废钢类型识别算法

Scrap Type Recognition Algorithm Based on Improved Inception-ResNet-V2

本研究提出了一种基于深度学习的废钢快速识别方法,提出的基于Inception-ResNet-V2的改进网络结构添加注意力机制模块经过微调得到SE-Inception-ResNet,并在此基础上采用学习率梯度更新策略自适应调节优化模型。采集了四种类型的废钢数据,然后将样本图像按80%训练集,20%验证集进行训练。后与ResNet152、InceptionV3比较了模型的性能。结果表明,SE-Inception-ResNet、InceptionV3和ResNet152网络的总体分类准确率分别为98.10%、97.48%、95.67%。SE-Inception-ResNet的分类精度最高,该模型在不同学习率情况下能快速梯度收敛。实验结果表明,所提出的改进卷积神经网络模型能够有效地对废钢类型进行识别。同时期望提高其迁移学习模型泛化性,可以为其他快速分类鉴定提供参考,并应用于其他工业或商业领域。

In this study,a deep learning-based method for rapid identification of scrap steel was proposed.The improved network structure based on Inception-Resnet-V2 added attention mechanism module was fine-tuned to obtain SE-Inception-ResNet,and on this basis,the learning rate gradient update strategy was used to adaptively adjust the optimization model.Then collected four types of scrap data,and then the sample images are trained with 80%training set and 20%validation set.The performance of the model was compared with ResNet152 and InceptionV3.The results show that the overall classification accuracy of SE-Inception-ResNet,InceptionV3,and ResNet152 are 98.10%,97.48%,and 95.67%,respectively.SE-Inception-ResNet has the highest classification accuracy,and the model can quickly converge with gradients under different learning rates.The experimental results show that the proposed improved convolutional neural network model can effectively identify the types of scrap steel.At the same time,it is expected to improve the generalization of its transfer learning model,which can provide a reference for other rapid classification and identification,and be applied to other industrial or commercial fields.