Study On Wood Identification Methods Of Pterocarpus Santalinus Based On Computer Vision

Due to the dramatic increase in illegal logging and trade,the wood resources are diminishing at an alarming rate,especially for rare wood species.To protect,manage and utilize wood resources in effective and rational ways,it is necessary to identify wood species accurately and efficiently.Moreover,the wide and indepth application of computer vision has driven the development of wood image intelligent recognition.However,at present,wood images are generally classified effectively by computer vision under certain conditions,which means that it is under the constrained image acquisition environment.So,it is difficult to meet the need for efficient wood identification.Moreover,it is difficult to extract effective macroscopic image features of wood with large variability,large intraclass difference and small inter-class difference.To accurately identify the macroscopic image of wood under unconstrained conditions,that is to reduce the requirements for image acquisition environment,this study aims to identify wood through deep learning and conventional machine learning after preprocessing with image enhancement and data augmentation,taking Pterocarpus santalinus,Pterocarpus tinctorius and Gluta sp.as the research objects as they share similar macro and micro structural characteristics.The main research work is as follows:Firstly,the wood image was preprocessed aiming at the problem that it was difficult to identify wood images under unconstrained conditions.On the one hand,image super-resolution reconstruction method was used to improve the quality of degraded images,so that more image details could be obtained.On the other hand,the data augmentation method was used to expand the wood image database,so as to increase the diversity of image sets during training model.Secondly,the macro image recognition models of the wood longitudinal section were constructed based on conventional machine learning.The image feature parameters,extracted by the gray level co-occurrence matrix,wavelet transform,histogram of oriented gradient,local binary pattern and visual bag of words,were fed into the classifiers which included linear discriminant analysis,K-nearest neighbor,support vector machine and artificial neural network.Furthermore,the prediction performance of the recognition models constructed by various algorithms was analyzed,and then the separability of image feature parameters and the effectiveness of image preprocessing methods were further verified by the visualization method of T-distributed stochastic neighborhood embedding.Experimental results showed that the characteristic parameters extracted by the visual bag of words based on speeded up robust features(SURF)were the most discriminative,and the model trained by the artificial neural network had the strongest performance,with the recognition accuracy of 96.3%,88.9% and 74.1% in three different image sets.After image preprocessing,the recognition accuracy reached to 96.3%,94.4 % and 82.4%.Finally,the macro image recognition model of wood longitudinal section was constructed based on deep learning.Several convolutional neural network(CNN)models,including Alex Net,VGGNet,Google Net,Res Net and Bilinear CNN,were trained for wood recognition.The results demonstrated that the optimal model was Res Net50 with the recognition accuracy of 97.2%,73.2% and 56.5%in three different image sets.After image preprocessing,the recognition accuracy increased to 98.2%,97.2% and 81.5%.On the whole,the model trained by deep learning outperformed that by conventional machine learning.In addition,the process of extracting specific features from the shallow layer of CNN to abstract features from the deep layer was demonstrated by visualization of feature map,and the feature visualization by gradient-weighted class activation mapping(Grad CAM)indicated that there were differences in the wood image region of interest of CNN before and after image preprocessing.In summary,following image preprocessing,the Pterocarpus santalinus recognition model constructed by computer vision has strong generalization ability and robustness,and can identify the macro images of the wood longitudinal section under unconstraint conditions effectively.This study provides new ideas for accurate,efficient and nondestructive intelligent identification of wood,and it has high application value.