基于轻量化卷积神经网络的金属断口图像识别  

Metal fracture recognition based on lightweight convolutional neural network

作　　者：闫涵 卢伟[1] 吴玉虎 YAN Han;LU Wei;WU Yu-hu(College of Control Science and Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学控制科学与工程学院,辽宁大连116024

出　　处：《控制与决策》2024年第9期2858-2866,共9页Control and Decision

基　　金：国家自然科学基金项目(62073056,61876029);辽宁省应用基础研究计划项目(2023JH2/101300207);大连市重点领域创新团队项目(2021RT14);新疆维吾尔自治区科技重大专项项目(2022A01001)。

摘　　要：工业环境下金属断口图像识别是金属失效分析的重要一环,具有重要的研究意义.卷积神经网络(convolutional neural networks,CNN)已被证实在图像识别任务中是有效的,但是在工业环境下的金属断口图像识别仍然面临以下问题:1)金属断口图像具有较强的类内复杂性和类间相似性;2)现有CNN网络结构复杂,参数较多,很难部署在嵌入式设备上.针对上述问题,提出一种基于轻量化CNN的金属断口图像识别方法.首先,设计一种多特征融合的CNN模型结构来提升网络的特征提取能力,并给出一种混合剪枝算法对网络进行轻量化处理来降低算法复杂度;然后,将重要超参数搜索视为优化问题,利用贝叶斯优化(Bayesian optimization,BO)算法进行求解,实现模型设计和剪枝过程的自动化;接着,以金属断口图像数据集为例进行实验分析,实验结果表明所提出模型仅需3.82 M的参数量即可实现97.56%的识别精度;最后,将训练好的模型部署到Jetson Nano嵌入式平台上,验证了所提出算法在实际应用中的可行性.The recognition of metal fracture images in an industrial environment plays a pivotal role in the analysis of metal failures and carries substantial research significance.While convolutional neural networks have been proven effective in image recognition tasks,the recognition of metal fracture images in an industrial environment still encounters the following challenges.1)Metal fracture images exhibit strong intra-class complexity and inter-class similarity.2)Existing CNN structures are complex,with a large number of parameters,which makes deployment on embedded devices challenging.To address the aforementioned problems,this paper proposes a metal fracture image recognition method based on the lightweight CNN.First,a CNN model structure with multi-feature fusion is designed to enhance the network’s feature extraction capability.Second,a hybrid pruning algorithm is proposed to slim the network and reduce the complexity of the algorithm.Simultaneously,the search process for important hyperparameters is treated as an optimization problem,and the Bayesian optimization(BO)algorithm is utilized to solve it,thereby automating the model design and pruning process.The experimental results show that the proposed method requires only 3.82 million parameters to achieve 97.56%recognition accuracy.The deployment on the Jetson Nano embedded platform verifies the practical feasibility of the proposed method.

关 键 词：深度学习 图像识别 金属断口 轻量化网络 贝叶斯优化 Jetson Nano 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]