基于自适应反馈机制的小差异化图像纹理特征信息数据检索  

Retrieval of texture feature information data for small differentiated images based on adaptive feedback mechanism

作　　者：刘洋[1] 毛克明[2] LIU Yang;MAO Keming(Library,Northeastern University,Shenyang,Liaoning 110819,China;Software College,Northeastern University,Shenyang,Liaoning 110819,China)

机构地区：[1]东北大学图书馆,辽宁沈阳110819 [2]东北大学软件学院,辽宁沈阳110819

出　　处：《江苏大学学报（自然科学版）》2025年第1期73-81,共9页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助面上项目(61773104)。

摘　　要：针对小差异化图像纹理相似度和噪声等因素导致纹理特征挖掘效果较差的问题,设计一种自适应反馈结合局部二值机制的小差异化图像纹理特征挖掘方法.使用规范割策略将图像数据各点拟作节点,使用节点间的连接线权重计算2点的相似度,采用支持向量机训练图像属性参数分类图像属性,进一步归纳图像类别.运用跳跃连接方法传输图像数据,将数据引入卷积神经网络剔除图像噪声.将中心点像素值当作反馈因子,创建自适应反馈判定条件,利用局部二值模式实现小差异化图像纹理特征挖掘.在MATLAB平台进行试验,从卷积神经网络收敛性、图像频谱纹理单元数、平均准确率、图像数据匹配度等方面进行了分析,分析结果表明:随着迭代次数不断增加,精度损失逐渐降低,基本收敛到稳定值,达到了预期训练效果;所提出方法挖掘的图像频谱纹理单元数3800个以上,更贴合人眼视觉信息;平均准确率为0.87,准确率@1、准确率@5和准确率@10的平均值分别为0.90、0.84和0.85;挖掘耗时低于5 s,图像数据匹配度高于90.3%,验证了所提出方法可在图像纹理特征识别操作中发挥应有作用.To solve the problem of poor texture feature mining performance caused by texture similarity and noise in small differentiated images,the small differentiated image texture feature mining method was designed with combining adaptive feedback and local binary mechanism.The standard cut strategy was used to model each point in the image data as node,and the similarity between the two points was calculated by the weights of connecting lines among the nodes.The support vector machine was adopted to train image attribute parameters for classifying image attributes and image categories.Using the skip connection method to transmit image data,the data was introduced into convolutional neural network to remove image noise.Taking the pixel value of the center point as feedback factor,the adaptive feedback judgment conditions were created,and the local binary patterns were used to achieve small differential image texture feature mining.The experiments were conducted on the MATLAB platform to analyze the convergence of convolutional neural networks,the number of spectral texture units in images,the average accuracy and the image data matching.The analysis results show that as the number of iterations is increased,the accuracy loss is gradually decreased and converged to stable value,which achieves expected training effect.The proposed method can mine over 3800 spectral texture units in images,which are more in line with human visual information.The average accuracy is 0.87,and the average values of accuracy@1,accuracy@5 and accuracy@10 are 0.90,0.84 and 0.85,respectively.The mining time is less than 5 seconds,and the image data matching degree is higher than 90.3%,which verifies that the proposed method can play important role in image texture feature recognition operations.

关 键 词：小差异化图像 纹理特征 数据挖掘 自适应反馈 属性分类 跳跃连接 局部二值模式 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]