多层卷积神经网络在图像目标标注中的应用  

Application of Multi-Layer Convolutional Neural Networks in Image Object Labeling

作　　者：吴文波 杨耀宁 禹谢华 WU Wen-bo;YANG Yao-ning;YU Xie-hua(School of Computer Information Minnan Science and Technology College,Quanzhou Fujian 362332,China;Fujian Provincial University,Key Laboratory of Big Data and Artificial Intelligence,Quanzhou Fujian 362332,China;School of Architecture and Planning,Yunnan University,Kunming Yunnan 650500,China;Berlin Universitaet of Technische,Berlin 10623,Germany)

机构地区：[1]闽南科技学院计算机信息学院,福建泉州362332 [2]大数据与人工智能福建省高校重点实验室,福建泉州362332 [3]云南大学建筑与规划学院,云南昆明650500 [4]柏林工业大学,德国柏林10623

出　　处：《计算机仿真》2025年第2期243-247,共5页Computer Simulation

基　　金：福建省教改项目基金(FBJG20210123);大数据与人工智能福建省高校重点实验室(闽教科[2019]67);省级一流本科专业建设点(SJZY-2022-01);省级虚拟仿真实验教学一流课程(SJKC-2020-03);大数据与人工智能福建省高校重点实验室(GXKYSY201901);校级科研团队(MKKYTD202302);校级虚拟教研室(MKXNJYS-2023-02);福建省中青年教师教育科研项目(科技类)(JAT231173)。

摘　　要：图像中的目标具有不同形状、大小、姿态和视角,在复杂场景、模糊边界或目标重叠的情况下,标注的精度会受到限制,导致标注结果的准确性不高。为有效解决以上问题,提出多层卷积神经网络在图像目标标注中的应用研究。利用经过下采样优化后的非局部均值(non-local means,NLM)算法去噪图像,并通过拉普拉斯变换法和Sobel梯度法增强图像,提升图像质量。采用多层卷积神经网络作为图像目标标注算法,并在其中增添批标准化(batchnormalization,BN)层提升其性能,为每个图像标签设置最优阈值,通过融合最优阈值的多层卷积神经网络实现图像目标标注。实验结果表明,所提方法的图像目标标注中全类平均精度高、整体运行速度快。Generally,the target in the image has different shapes,sizes,postures and perspectives.In complex scenes,fuzzy boundaries or overlapping targets may affect the accuracy of annotation,resulting in inaccurate results.Therefore,an application of multi-layer convolutional neural networks in image annotation was proposed.Firstly,the non-local means(NLM)algorithm optimized by downsampling was used to denoise the image,and then the image quality was enhanced by the Laplacian transform method and Sobel gradient method.Moreover,a multi-layer convolutional neural network was used as the algorithm for image annotation.Meanwhile,a batch normalization(BN)layer was added to improve the performance of the algorithm.Finally,an optimal threshold was set for each image label,and the multi-layer convolutional neural network with the optimal threshold was used to annotate the image target.Experimental results show that the proposed method has high average precision and fast overall running speed in image annotation.

关 键 词：多层卷积神经网络 图像目标标注 非局部均值去噪 非锐化掩蔽增强 批标准化层 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]