Study Of Wood Microscopic Image Identification Based On Transfer Learning Multi-Fusion Model

There are a great variety of wood whose materials,functions and business value vary greatly.Accurate classification and identification of wood is an important way to use it rationally.The traditional computer-aided classification and wood identification technologies based on chemical composition all have obvious limitations,which restrict the application and promotion in the field of wood identification.Therefore,it is of great practical significance to study the identification methods based on deep learning and computer vision identifying wood species automatically,efficiently and accurately.The rich wood structure features which provide effectively feature information for wood identification can be reflected by microscopic images of wood.Automatic classification and identification of wood microscopic images by computers has important research value.Wood microscopic image is a kind of fine-grained image.In some cases,due to the small differences between wood species,large intra-species variability,and the small number of image samples,the difficulty of classification and identification of wood microscopic images is increased.In order to solve these problems.Firstly,the combined expansion strategy is used to amplify the number of sample images and to establish a wood microscopic image data set.Secondly,a wood microscopic image classification model based on the decomposition-aggregation module and a wood microscopic image detection model on multi-fusion is constructed.Thirdly,A classification and identification method of wood microscopic images based on the transfer learning fusion model is proposed to achieve the purpose of improving the accuracy of wood microscopic image classification and identification.The main research content and results of this paper are as follows:(1)To study the amplification method of small sample data set and the construction of wood microscopic image sample library and data set.Taking account of the content and objectives in this paper,each type of wood sample used for wood classification and identification model should have a sufficient number to meet the needs of training for the deep learning model reasonably.This paper selects representative wood samples from 8 genera and 56 species,and uses Nikon Eclipse Ni-U microscope with Nikon DS-Ri2 camera(or Nikon 80i microscope with Nikon DS-Ril-U3 camera)to collect the images and NIS-Elements microscopic image software to process them.The magnification of wood microscopic images is 4 times,10 times,20 times and 40 times,respectively.The images of each wood species come from 10 slices taken from different material parts,and 3928 images containing sample library and data set of wood microscopic images for 56 species of wood are established.In order to reduce the workload of wood microscopic image acquisition and avoid the problem of poor performance of small sample data sets in network training and testing,an online wood microscopic image amplification method based on the combined expansion strategy MixupCES is proposed.To generate more abundant sample data sets online to ensure that the model obtains more condensed wood species information for more adaptability and generalization ability and to increase the detectability for strange sample,the image geometric transformation and Mixup data expansion technology are integrated.The experimental results show that the method in this paper makes the network more specific to the learning of wood microscopic image inter-species similarity and intra-species diversity,which is better than traditional data expansion and Mixup data expansion methods and provides data support for the study of wood identification in this paper.(2)In order to solve the problem that the convolution structure of the classic convolutional neural network is difficult to obtain the overall characteristics and long-term dependence of the wood microscopic image in the same layer network structure,a wood microscopic image based on the decomposition and aggregation network model is proposed.Aiming at the problem of ignoring the correlation among the features in non-local neural network,a ResNet-M network model based on decomposition and aggregation is designed and constructed.In this way that the decomposition and aggregation module is added to each block of the ResNet the decomposition structure is to better captures the importance of different features in each part of the channel and space and the aggregation structure is used to promote the integration of the enhanced features of the two channels.And capturing ability of the overall characteristics and long-term dependence relation of image and identifying ability of tiny difference among wood species are promoted.The experimental results show that all indicators of the method in this paper are better than the three methods of VGGNet16,InceptionV3 and ResNet50,and the identification ability on the microscopic images of similar species is enhanced,and it can distinguish the wood of the same genus which is easy to be confused.(3)Due to the fact that the original features can be extracted and selected on the basis of classification for target detection tasks,and compared with simple classification,the identification results and the characteristics of confidence can be displayed more intuitively.A wood microscopic image identification method based on multiple fusion detection models is proposed to improve the traditional Faster R-CNN algorithm for wood microscopy.In this model we use the ResNet50-M network to fix the first two convolutional blocks under the framework of Faster R-CNN,and use the fine-tuned three convolutional blocks as a transfer learning strategy to extract wood microscopic image features,then we introduce SPP,CBAM,FPN and softer-NMS model to provide feature maps of wood microscopic images of different scales for the RPN network to generate high-quality candidate frames.Finally,we construct a multi-fusion detection model TLMFNet based on transfer learning,and verify the best component composition of TLMFNet through ablation experiments.The experimental results show that the method in this paper analyzes and integrates the feature information of wood microscopic images from a relatively global perspective,and increases the weight of features related to wood species identification tasks.It is better than model methods such as YOLOv3,Faster R-CNN and Faster R-CNN w FPN.It can better identify different species of wood in the same genus.(4)In view of the invisible problem of the internal algorithm of deep learning,we visualize the wood microscopic image features captured by the TLMFNet backbone network ResNet50-M,then analyze the method and process of TLMFNet extracting wood microscopic image features,and finally we visualize the detection results to evaluate the performance of TLMFNet in wood microscopic image inspection tasks.The experimental results show that the transfer learning strategy of the TLMFNet model in this paper is reasonable.Compared with the ResNet-M model that simply improves the classification ability,the multi-fusion improvement strategy of the detection model TLMFNet is an overall improvement in the feature extraction ability,and it also plays a role in detection and classification.The recognition precision,positioning accuracy and confidence of wood microscopic images are also improved,which shows that the multi-fusion improvement strategy in this paper is necessary and effective for obtaining discriminative information of wood species.In summary,the wood microscopic image identification method based on the transfer learning fusion model can improve the efficiency of wood microscopic image classification and identification.It can be used in the identification and rational utilization of wood species and has certain research value and guiding significance.