融合频谱变换的板材纹理缺陷分类

摘要

In order to automatically recognize and classify of wood surface texture and defects, we proposed a rapid and collabo-rative classification method, which fused spectrum transforms of wavelet, curvelet, and dual-tree complex wavelet.Initial-ly , we extracted 14 features of wavelet, 16 features of curvelet, and 38 features of dual-tree complex wavelet.We fused three characteristics and the whole image of the standard deviation and entropy, and used particle swarm algorithm to select 24 key features.Finally, we used BP neural network as a classifier to take the simulation experiment with 300 images of 5 categories, including turbulence, parabolic lines,straight lines , slipknot and knot.The classification correct rate of wave-let transform, curvelet transform, dual-tree complex wavelet transform and feature fusion were 80.0%, 81.1%, 84.2%and 88.0%, respectively,and the classification time were 0.018, 0.503, 0.021, and 0.325s , respectively.The feature fusion can effectively select the feature of wavelet, curvelet and dual-tree complex wavelet, and improve the speed and accuracy of classification.%为了实现板材表面纹理和缺陷的自动分类识别，提出一种融合小波、曲波和双树复小波3种频谱变换的板材表面纹理和缺陷的快速协同分类方法。分别提取小波变换的14个特征、曲波变换的16个特征和双树复小波变换的38个特征；融合三者特征以及整幅图像的标准差和熵，采用粒子群算法优选出24个关键特征；运用BP神经网络作为分类器，针对乱纹、抛物纹、直纹、活结和死结5种类别的300幅图像进行仿真实验，基于小波变换、曲波变换、双树复小波变换与特征融合方法的平均分类正确率分别为80．0％、81．1％、84．2％、88．0％，分类平均时间分别为00．18、0．503、0．021、0．325 s。实验结果表明，特征融合方法实现了对小波特征、曲波特征和双树复小波特征的有效选择，提高了分类的速度和精度。