Intelligent identification method for main anatomical features of coniferous tree based on deep learning

Abstract

The invention discloses a coniferous tree main anatomical feature intelligent identification method based on deep learning, and the method comprises the steps: carrying out the image preprocessing and image enhancement technologies; the method comprises the following steps: establishing a high-quality and sufficient-quantity coniferous tree three-section image database, and establishing a suitable deep learning model for five anatomical features of morning and evening wood transformation degrees, transverse resin traces, axial resin traces, ray cells and cross field holes; on this basis, an image processing algorithm is combined to realize automatic identification of wood anatomical features, an optimization method is selected to improve the fitting ability of a model to a target task, then the feature prediction accuracy is improved, a decision tree algorithm is used to carry out classification prediction on coniferous trees, the error rate of the pinaceae wood anatomical features identified by the method is less than 5%, and the method is suitable for popularization and application. The identification features are all wood universal features, and the limitation is small; the accuracy rate of predicting tree species names through wood anatomical features is higher than 90%, and the actual use requirement can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of wood anatomy, and in particular to an intelligent recognition method for main anatomical features of coniferous wood based on deep learning. Background Art

[0002] Different tree species have different wood structures, physical and chemical properties, and different application scenarios and usage values. Therefore, it is of practical significance to quickly, efficiently and accurately realize the intelligent recognition of wood anatomical characteristics and explore the application of recognition methods in the study of wood anatomical structure and tree species identification.

[0003] With the rapid development of deep learning algorithms, image classification, semantic segmentation, target detection and other technologies in computer vision have become more mature, and the application scenarios are constantly expanding. In the process of development, applying computer vision technology to wood feature recognition has become a popular development direction. At present, in the field of wood science, the research using computer vision technology mainly includes the use of gray-level co-occurrence matrix or convolutional neural network to extract abstract feature information from wood images for wood recognition. However, the microstructure and composition of wood are not discussed in this process. Therefore, the established prediction model can often only identify tree species within a small range. For the identification of tree species outside the range, a new model needs to be established, and it is difficult to form a complete set of recognition evaluation standards. In addition, as a natural biomass material, wood has intraspecific variation. When dealing with the prediction task of wood images of unknown origin, the above classification model has poor generalization and cannot guarantee reliable accuracy. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] Therefore, the purpose of the present invention is to provide a method for intelligently identifying the main anatomical features of coniferous wood based on deep learning, and to achieve a method for classifying and predicting tree species. The method is simple, efficient, accurate and has good versatility.

[0006] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0007] The intelligent recognition method of the main anatomical characteristics of coniferous wood based on deep learning includes:

[0008] S1. Establish a high-quality and sufficient three-section image database of coniferous wood through image preprocessing and image enhancement technology;

[0009] S2. Establish a suitable deep learning model for the five anatomical features of early-late wood transition, transverse resin canals, axial resin canals, ray cells, and cross-field pits; on this basis, combine image processing algorithms to realize automatic recognition of wood anatomical features;

[0010] S3. Use optimization methods to improve the model's ability to fit the target task, thereby improving feature prediction accuracy;

[0011] S4. Use decision tree algorithm to classify and predict coniferous timber.

[0012] As a preferred solution of the method for intelligently identifying the main anatomical features of coniferous wood based on deep learning described in the present invention, in step S1, the specific steps of establishing a high-quality and sufficient three-section image database of coniferous wood by image preprocessing and image enhancement technology are as follows:

[0013] Use a high-precision microscope and image acquisition system to perform microscopic imaging on wood samples of different tree species to obtain microscopic structural images of wood; or use a high-quality wood image database to obtain microscopic structural images of wood;

[0014] The collected images are preprocessed including grayscale conversion, size normalization, and histogram equalization to improve image quality and prepare for feature extraction;

[0015] The dataset is then expanded using rotation, cropping, color adjustment, and scaling image augmentation methods.

[0016] As a preferred solution of the method for intelligently identifying the main anatomical features of coniferous wood based on deep learning described in the present invention, in step S2, a suitable deep learning model is established for the five anatomical features of earlywood to latewood transition, transverse resin canals, axial resin canals, ray cells, and cross-field pits; on this basis, combined with the image processing algorithm, the automatic recognition of the anatomical features of wood is realized specifically as follows:

[0017] For the early and late wood transition degree classification task, the input image size is 224×336, divided into training set and prediction set in a ratio of 9:1, and the Resnet50 model is selected for training;

[0018] For the resin channel feature recognition task, the input image size is 512×768, which is divided into a training set and a prediction set in a ratio of 9:1, and the U-Net model is selected for training;

[0019] For the cross-field pit type detection task, the input image size is 416×416, divided into a training set and a prediction set in a ratio of 9:1, and the YOLOv4 model is selected for training;

[0020] For the ray size measurement task, the input image size is 512×768, which is divided into a training set and a prediction set in a ratio of 9:1, and U-Net is selected for semantic segmentation training;

[0021] For the ray size measurement task, the semantically segmented image is preprocessed using the flood filling method, and the ray height is calculated using the minimum enclosing circle algorithm;

[0022] For the X-ray cell counting task, the input image size is 128×512, which is divided into a training set and a prediction set in a ratio of 9:1, and the YOLOv4-tiny model is selected for training.

[0023] As a preferred solution of the method for intelligently identifying the main anatomical features of coniferous wood based on deep learning described in the present invention, in step S4, the step of using the decision tree algorithm to classify and predict coniferous wood is specifically as follows: performing statistical work on the anatomical features of wood identified in the image and establishing a data set;

[0024] Among them, the selected characteristics specifically include: early-latewood transition type, presence of transverse resin canals, presence of axial resin canals, average number of ray cells, and cross-field pit type;

[0025] The dataset is labeled and divided into a training set and a prediction set in a ratio of 7:3, and the ID3 decision tree is used for training.

[0026] Compared with the prior art, the present invention has the following beneficial effects: the present invention has the following outstanding and significant advantages: unlike other methods that use deep learning models to directly classify wood images, this method first extracts the anatomical features in the wood slice image, and then realizes wood recognition through a certain matching algorithm. This method greatly increases the versatility of the recognition algorithm. After further expanding the matching relationship between anatomical features and tree species, it can be widely used in all coniferous wood recognition tasks. In terms of recognition effect, among the anatomical features selected by this method, the early and late wood transition type, the presence or absence of transverse and axial resin canals, the size and number of ray cells, and the type of cross-field pits all show excellent extraction and recognition effects. The comprehensive success rate of the four indicators is above 90% after optimization; after comprehensively considering the four indicators using the decision tree algorithm, the accuracy of the tree species classification results is also higher than 90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0028] Figure 1 A flow chart of the method for intelligently identifying the main anatomical features of coniferous wood based on deep learning provided by the present invention;

[0029] Figure 2 Part of the result diagram of the image processing of the embodiment provided by the present invention;

[0030] Figure 3 Part of the result diagram of the expansion of the data set of the embodiment provided by the present invention;

[0031] Figure 4 The model training result of the early and late wood transformation degree classification task provided by the embodiment of the present invention;

[0032] Figure 5 The horizontal model training results of the resin channel recognition task provided by the embodiment of the present invention;

[0033] Figure 6 The longitudinal model training results of the resin channel recognition task provided in the embodiment of the present invention;

[0034] Figure 7 The prediction result of the cross-field pit recognition task provided by the embodiment of the present invention;

[0035] Figure 8 The detection result of the ray size measurement task provided by the embodiment of the present invention;

[0036] Fig. 9 The test result of the ray cell counting task provided in the embodiment of the present invention;

[0037] Fig.10 The resin channel recognition process of the embodiment provided by the present invention is a network structure diagram before and after the U-Net network is adaptively adjusted using transfer learning;

[0038] Fig.11 The comprehensive error rate and mean deviation of cell counting after the attention mechanism of the embodiment provided by the present invention is introduced;

[0039] Fig.12 A confusion matrix diagram of tree species prediction implemented by a decision tree model in an embodiment provided by the present invention. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0041] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0042] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0043] The present invention provides a method for intelligently identifying the main anatomical features of coniferous wood based on deep learning, and a method for realizing classification prediction of tree species. The method is simple, efficient, accurate and has good versatility.

[0044] like Figure 1 As shown in the figure, the intelligent recognition method of the main anatomical characteristics of coniferous wood based on deep learning has the following specific steps:

[0045] Through the wood image network database, the Nanjing Forestry University Wood Specimen Museum independently made slices and collected microscopic images of coniferous wood.

[0046] More than 2,000 images that met the research requirements were collected from the online database, covering a total of 47 tree species.

[0047] More than 600 wood slices were selected from the Wood Specimen Museum of Nanjing Forestry University, covering a total of 60 tree species.

[0048] The images were collected using an Olympus BX51 biological optical microscope and an Olympus DP70 microscope digital camera. A total of more than 1,800 images were collected.

[0049] The wood microscopic images are preprocessed and the dataset is expanded using image enhancement techniques.

[0050] Image preprocessing mainly includes three methods: image grayscale, size normalization, and histogram equalization. The factors that may interfere with the identification of wood anatomical features are eliminated to retain more effective feature information. The processed image results are as follows: Figure 2 shown.

[0051] Image enhancement technology mainly increases the total number of data set samples through geometric transformation, chromaticity adjustment and other means to improve the generalization of deep learning models. Some results of data set expansion are shown in the figure below. Figure 3 shown.

[0052] Pixels in random areas of the image are removed to reduce image incompleteness caused by occlusions such as impurities and noise.

[0053] After expansion, more than 20,000 images of wood microscopic features were collected.

[0054] For the early and late wood transition degree classification task, the Resnet50 model is selected for training. The model training results of the early and late wood transition degree classification task are as follows: Figure 4 shown.

[0055] The input image size is 224×336, which is divided into training set and prediction in a ratio of 9:1. The softmax (cross entropy loss) classification algorithm is used, and the training epoch is 200.

[0056] When the number of iterations reaches about 125, the loss rate tends to be flat, the training set loss values ​​of the model are all stable below 0.2, and the test set loss values ​​are all below 0.05 and smaller than the training set loss values, indicating that the model has not overfitted after the number of iterations reaches 200.

[0057] For the resin channel feature recognition task, the U-Net model is selected for training. The horizontal and axial model training results of the resin channel recognition task are as follows: Figure 5 and Figure 6 shown.

[0058] The input image size is 512×768, which is divided into a training set and a prediction set in a ratio of 9:1. The Adam optimizer is used, and the training generation is 200. A hybrid strategy is adopted for resin channel determination, and misjudged images containing small areas of suspected resin channels are screened out.

[0059] When the number of iterations reaches about 50, the loss rate tends to be flat, and the training set loss value of the model is stable below 0.2.

[0060] For the cross-field pit type detection task, the YOLOv4 model is selected for training. The prediction results of the cross-field pit recognition task are as follows: Figure 7 shown.

[0061] The input image size is 416×416, which is divided into training set and prediction set in a ratio of 9:1. The Adam optimizer is used and the training epoch is 100.

[0062] When the number of iterations reaches about 10, the loss rate tends to be flat, and the loss value of the YOLOv4 model is below 0.05.

[0063] For the ray size recognition task, U-Net is selected for semantic segmentation training. The detection results of the ray size measurement task are as follows: Figure 8 shown.

[0064] The input image size is 512×768, which is divided into training set and prediction set in a ratio of 9:1. The softmax (cross entropy loss) classification algorithm is used, and the training epoch is 200.

[0065] When the number of iterations reaches 10, the Loss value gradually stabilizes below 0.1, and the surface model can correctly extract the wood ray features.

[0066] The semantically segmented image is preprocessed using the flood filling method to remove incomplete rays at the edge of the image, and the minimum enclosing circle algorithm is used to calculate the ray height.

[0067] For the ray cell counting task, the YOLOv4-tiny model is selected. The detection results of the ray cell counting task are as follows Fig. 9 shown.

[0068] The input image is 128×512 pixels, the stochastic gradient descent optimizer is selected, the ratio of training set to validation set is 9:1, and training is performed for 200 generations.

[0069] When the number of iterations reaches about 20, the loss rate tends to be flat, and the loss value of the YOLOv4 model is below 0.1, indicating that the surface model is suitable for the ray cell counting task.

[0070] The deep learning model optimization method and effect evaluation described in the present invention include the following contents:

[0071] The resin channel feature recognition model is optimized using transfer learning. The resin channel recognition process uses transfer learning to make adaptive adjustments to the U-Net network before and after the network structure diagram is shown below. Fig.10 shown.

[0072] The main components of wood microscopic slice images are simple and repetitive cell structures, which are very similar to medical images. Therefore, we try to use the pre-trained weights of the medical data set in the VOC database as the initial training weights to test the prediction effect of the improved model.

[0073] During the training process, the backbone feature extraction network of the model is frozen to achieve transfer learning. In order to ensure the adaptability of the U-Net network to the ImageNet medical imaging dataset, the original U-Net network is adaptively adjusted and modified according to the VGG16 network (such as Fig.10 The size of the feature layer of the main feature extraction part and the enhanced feature extraction part was changed to make the final feature layer the same size as the original input image. The mIoU value was selected as the U-Net performance evaluation indicator. The results showed that the machine misjudged other cells as resin ducts in many cases, and the medical data set was higher than the ordinary data set, while the case of missing resin ducts was almost non-existent, which was speculated to be due to overfitting of the model.

[0074] Introducing the ECANet attention mechanism into the YOLOv4-tiny model for the cell counting task.

[0075] As another implementation of the channel attention mechanism, ECANet removes the fully connected layer in the original SENet and uses a 1×1 convolution layer directly after the global average pooling layer to complete the cross-channel information interaction. The size of the convolution kernel changes adaptively through a function, so that the layer with a larger number of channels can interact more across channels, thus avoiding dimensionality reduction. In this example, only a small number of parameters are involved to achieve good results, and the mean deviation is below 0.03. The comprehensive error rate and mean deviation of cell counting after the introduction of the attention mechanism are shown in the following figure. Fig.11 shown.

[0076] Classification of tree species based on decision tree algorithm

[0077] By establishing an ID3 decision tree model, the tree species name is predicted based on the coniferous wood characteristic information.

[0078] We selected 10 genera of tree species in the Pinaceae family, including Abies, Cathaya, Chinese Fir, Larix, Picea, Golden Larch, Pseudotsuga, Tsuga, Pinus, and Cedrus, as the identification test objects. We collected more than 400 slice images of the above 10 genera of tree species in the Pinaceae family, and conducted feature statistics. We selected five features, including the type of early and late wood transition, the presence or absence of transverse resin canals, the presence or absence of axial resin canals, the average number of ray cells, and the type of cross-field pits, as the classification basis.

[0079] The early and late wood transition degree is taken as feature 1, where the slow change type is marked as 0 and the sudden change type is marked as 1; the presence of axial resin canals and transverse resin canals is taken as feature 2 and feature 3, where the absence of resin canals is marked as 0 and the presence of resin canals is marked as 1; the average value of the number of ray cells is taken as feature 4, and its average value is used as the marking information; the cross-field pit type is taken as feature 5, where the spruce type is marked as 0, the cypress type is marked as 1, the fir type is marked as 2, the pine type is marked as 3, and the window-shaped type is marked as 4. The labeled data is trained using the ID3 decision tree, and the ratio of the training set to the prediction set is 7:3.

[0080] The confusion matrix of tree species prediction results is as follows Fig.12 As shown in the figure, it can be seen that a total of 12 samples out of the 116 samples used for prediction were incorrectly predicted, and the prediction accuracy was about 90%. Among them, the samples with the most prediction errors appeared in the case where the spruce species were mistakenly judged as the silver fir species. However, this is due to the limited number of selected features. The five groups of features selected for the tree species of the two genera are very similar in themselves, making it difficult to predict correctly. This can be solved later by expanding the number of selected feature types.

[0081] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An intelligent identification method for the main anatomical features of coniferous wood based on deep learning, characterized in that: include: S1. Establish a high-quality and sufficient three-section image database of coniferous wood through image preprocessing and image enhancement technology; S2. Establish a suitable deep learning model for the five anatomical features of early-late wood transition, transverse resin canals, axial resin canals, ray cells, and cross-field pits; on this basis, combine image processing algorithms to realize automatic recognition of wood anatomical features; S3. Use optimization methods to improve the model's ability to fit the target task, thereby improving feature prediction accuracy; S4. Use decision tree algorithm to classify and predict coniferous timber.

2. The method for intelligently identifying the main anatomical features of coniferous wood based on deep learning according to claim 1, characterized in that: In step S1, the specific steps of establishing a high-quality and sufficient three-section image database of coniferous wood by image preprocessing and image enhancement technology are as follows: Use a high-precision microscope and image acquisition system to perform microscopic imaging on wood samples of different tree species to obtain microscopic structural images of wood; or use a high-quality wood image database to obtain microscopic structural images of wood; The collected images are preprocessed including grayscale conversion, size normalization, and histogram equalization to improve image quality and prepare for feature extraction; The dataset is then expanded using rotation, cropping, color adjustment, and scaling image augmentation methods.

3. The method for intelligently identifying the main anatomical features of coniferous wood based on deep learning according to claim 1, characterized in that: In step S2, a suitable deep learning model is established for the five anatomical features of early-late wood transition, transverse resin canals, axial resin canals, ray cells, and cross-field pits; on this basis, an image processing algorithm is combined to realize the automatic recognition of wood anatomical features, specifically: For the early and late wood transition degree classification task, the input image size is 224×336, divided into training set and prediction set in a ratio of 9:1, and the Resnet50 model is selected for training; For the resin channel feature recognition task, the input image size is 512×768, which is divided into a training set and a prediction set in a ratio of 9:1, and the U-Net model is selected for training; For the cross-field pit type detection task, the input image size is 416×416, divided into a training set and a prediction set in a ratio of 9:1, and the YOLOv4 model is selected for training; For the ray size measurement task, the input image size is 512×768, which is divided into a training set and a prediction set in a ratio of 9:1, and U-Net is selected for semantic segmentation training; For the ray size measurement task, the semantically segmented image is preprocessed using the flood filling method, and the ray height is calculated using the minimum enclosing circle algorithm; For the X-ray cell counting task, the input image size is 128×512, which is divided into a training set and a prediction set in a ratio of 9:1, and the YOLOv4-tiny model is selected for training.

4. The method for intelligently identifying the main anatomical features of coniferous wood based on deep learning according to claim 1, characterized in that: In step S3, the optimization method is specifically selected as follows: For deep learning models, select the Adam optimizer; Make adaptive adjustments to the U-Net network and modify the original U-Net network based on the VGG16 network to achieve transfer learning; A hybrid strategy is adopted for resin channel determination to screen out misjudged images containing small areas of suspected resin channels; For the cross-field pit type detection task, the Grad-CAM method is used for visualization and weighted summation through the ReLU function to obtain the heat map result of the detection image; For the ray cell counting task, the ECANet attention mechanism is introduced.

5. The method for intelligently identifying the main anatomical features of coniferous wood based on deep learning according to claim 1, characterized in that: In step S4, the steps of using the decision tree algorithm to classify and predict coniferous wood are specifically as follows: performing statistical work through the anatomical features of the wood identified in the image and establishing a data set; Among them, the selected characteristics specifically include: early-latewood transition type, presence of transverse resin canals, presence of axial resin canals, average number of ray cells, and cross-field pit type; The dataset is labeled and divided into a training set and a prediction set in a ratio of 7:3, and the ID3 decision tree is used for training.

Images