基于卷积神经网络的汽车图像损坏检测  

Vehicle image damage detection based on Convolutional Neural Networks

作　　者：王纯杰 易铭瑒 谭佳伟 WANG Chunjie;YI Mingyang;TAN Jiawei(School of Mathematics&Statistics,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]长春工业大学数学与统计学院,吉林长春130012

出　　处：《长春工业大学学报》2024年第3期193-198,F0003,共7页Journal of Changchun University of Technology

基　　金：吉林省科技厅重大科技专项(20210301038GX;20220301031GX)。

摘　　要：探讨卷积神经网络(CNN)在汽车图像中损坏识别的应用,通过收集包含不同损坏程度的汽车图像数据集构建一个CNN模型,利用准确率等关键性能指标对该模型进行深入评估。实验对比多种损失函数对模型性能的影响,分析表明,采用稀疏类别交叉熵损失函数的CNN模型在性能表现上较为突出,其准确率达到97%。这一发现证明,稀疏类别交叉熵在提升模型准确性方面的显著优势,本研究为利用CNN在汽车图像中实现损坏识别提供有力支持。This study aims toexplore the application of Convolutional Neural Networks(CNN)in the identification of damages in car images.By collecting a dataset of car images with varying degrees of damage,a CNN model was constructed and subjected tOan in-depth evaluation using key performance indicators such as accuracy.The experiment compared the effects of various loss functions on model performance,and the analysis revealed that the CNN model employing Sparse Categorical Cross-Entropy loss function exhibited superior performance,achieving an accuracy rate of 97%.This finding demonstrates the significant advantage of Sparse Categorical Cross-Entropy in enhancing model accuracy.Overall,this research provides strong support for the use of CNN in the identification of damages in car images.

关 键 词：卷积神经网络 深度学习 分类模型 汽车图像识别 

分 类 号：O212[理学—概率论与数理统计] O212.8[理学—数学]