红外图像识别的Transformer频域可学习性方法  

A Transformer Frequency Domain Learnability Method for Infrared Image Recognition

作　　者：赖光明 张倬实 郭鑫平 汪敏 LAI Guangming;ZHANG Zhuoshi;GUO Xinping;WANG Min(School of Electrical Information Southwest Petroleum University,Chengdu,610000 China)

机构地区：[1]西南石油大学电气信息学院,成都610000

出　　处：《电光与控制》2023年第8期13-18,共6页Electronics Optics & Control

基　　金：国家自然科学基金(62006200);四川省科技计划支持项目(2022YFG0179);油气藏地质及开发工程国家重点实验室(成都理工大学)项目(PLC20211104);中国石油-西南石油大学创新联合体科技合作项目(2020CX020000);四川省科技创新苗子工程(2022032)。

摘　　要：伴随工业自动化的发展,红外图像识别技术更多地应用于自动化生产领域。红外图像存在噪点多、图像质量差、色彩信息缺失等特点。针对上述特点,提出一种从红外图像频率信息出发,对红外图像进行识别的检测方法——红外图像频域检测方法(IFDM)。首先,有别于传统图像处理,该方法从频域角度出发,通过离散傅里叶变换,将图像信息变换到频域,有利于更好地把握红外图像特有的结构特征;其次,在频域进行可学习的频率信息筛选,增强了模型的特征提取能力;最后,引入了Transformer结构,相较于CNN结构,该结构能更好地整合图像中的全局信息。通过在3个特有的红外图像数据集上进行测试,与其他算法在准确率以及模型收敛速度层面进行比较,验证了该方法的可行性。With the development of industrial automation infrared image recognition technology is applied more frequently to the field of automated production.Infrared images are characterized by high noise poor image quality and lack of color information.In view of the above characteristics a detection method Infrared Image Frequency Domain Detection Method(IFDM)is proposed based on infrared image frequency information to identify infrared images.Firstly different from traditional image processing this method starts from the frequency domain and transforms the image information into the frequency domain through discrete Fourier transform which is beneficial to better grasp the unique structural features of infrared images.Secondly the learnable screening of frequency information in the frequency domain enhances the feature extraction capability of the model.Finally the Transformer structure is introduced which can better fuse the global information in the image than the CNN structure.Three unique infrared image datasets are used to verify the feasibility of the method in comparison with other algorithms in terms of accuracy and model convergence rate.

关 键 词：红外图像识别 机器学习 TRANSFORMER 频率深度学习 

分 类 号：TN219[电子电信—物理电子学]