Statistical feature representations for automatic wood defects recognition research and applications

Abstract

In this paper, we introduce the non-negative matrix factorization (NMF) to decompose the wood images and structure the feature spaces. Local binary pattern (LBP) is used to extract the original spatial local structure features, such as curly edges, etc. and they have better luminance adaptability. Simultaneously, dual-tree complex wavelet transform (DTCWT) is used to extract the energy based statistical features from different directions and frequencies and they can maintain better time-frequency localized characteristics and finite data redundancy. We integrate the features together to choose the proper features to describe the discrepancies between sound woods and defects and then propose an automatic detection system for wood defects recognition. After many cross experiments, we received a better identification rate of more than 90% with good research values and potential applications.