Wood species recognition method based on multi-feature fusion

Abstract

The present invention discloses a method for identifying wood species based on multi-feature fusion, and the method includes image acquisition, image denoising, image segmentation and wood species identification. The present invention belongs to the technical field of tree species identification, and specifically refers to a method for identifying wood species based on multi-feature fusion. The scheme introduces a gain coefficient into a generalized Anscombe transformation to more finely control the noise weight change of different pixels of wood species images; introduces dynamic threshold screening, and obtains wood species denoising results through multi-scale fusion, which greatly improves the accuracy of the final wood species identification; introduces a multi-scale feature aggregation module, and constructs a multi-level fusion block based on dilated convolution and dynamic convolution kernel selection mechanism, which helps to capture the annual rings and large texture structures in the wood; defines the segmentation loss function through weight addition, and improves the segmentation accuracy of wood species images with imbalanced categories; thereby ensuring the accuracy of the final wood species identification.

Description

Technical Field

[0001] The present invention relates to the technical field of tree species identification, and in particular to a wood species identification method based on multi-feature fusion. Background Art

[0002] Timber species identification refers to the method of analyzing and processing the characteristics of wood samples to determine the species to which they belong. This method is widely used in the fields of wood science, ecology, forestry management and archaeology. However, general timber species identification methods are unable to adapt to the noise changes in different areas of complex timber species images, and there is a loss of edges and details when denoising timber species images, which leads to low final recognition accuracy. General timber species identification methods have the problem of insufficient information recovery, resulting in insufficient detail representation of timber species images, inability to take into account global and local information, ignoring subtle texture and color changes of wood, and difficulty in dealing with the imbalance of timber species categories. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a wood species recognition method based on multi-feature fusion. In view of the fact that general wood species recognition methods are unable to adapt to the noise changes in different areas of complex wood species images, and there is a loss of edges and details when denoising wood species images, which leads to low final recognition accuracy, this scheme introduces a gain coefficient into the generalized Anscombe transformation to more finely control the noise weight changes of different pixels in the wood species image; introduces dynamic threshold screening to compress the noise components below the threshold in the wood species image, while avoiding excessive smoothing of important information, and obtains the wood species denoising result through multi-scale fusion, which greatly improves the accuracy of the final wood species recognition; in view of the fact that general wood species recognition methods are unable to adapt to the noise changes in different areas of complex wood species images, and there is a loss of edges and details when denoising wood species images, which leads to low final recognition accuracy, this scheme introduces a gain coefficient into the generalized Anscombe transformation to more finely control the noise weight changes of different pixels in the wood species image; introduces dynamic threshold screening to compress the noise components below the threshold in the wood species image, while avoiding excessive smoothing of important information, and obtains the wood species denoising result through multi-scale fusion, which greatly improves the final accuracy of wood species recognition; The method is not fine enough in restoring information, resulting in insufficient details of wood species images, failure to take into account global and local information, ignoring subtle texture and color changes of wood, and difficulty in dealing with the problem of imbalanced wood species categories. This scheme introduces a multi-scale feature aggregation module to ensure the delicate restoration of wood species image information. Through a cross-level feature fusion strategy, it can retain deep semantic information while restoring shallow details when processing annual rings and fine lines in wood images; construct a multi-level fusion block based on dilated convolution and dynamic convolution kernel selection mechanism to help capture annual rings and large texture structures in wood; define the segmentation loss function through weight addition to improve the segmentation accuracy of wood species images with imbalanced categories; and thus ensure the accuracy of the final wood species recognition.

[0004] The technical solution adopted by the present invention is as follows: The method for identifying wood species based on multi-feature fusion provided by the present invention comprises the following steps:

[0005] Step S1: image acquisition;

[0006] Step S2: image denoising;

[0007] Step S3: image segmentation;

[0008] Step S4: Identify wood species.

[0009] Furthermore, in step S1, the image acquisition is to acquire historical wood species images; the acquired historical wood species images are size-normalized, and the images are classified and labeled, and the labeled content is the specific category of the wood species.

[0010] Furthermore, in step S2, the image denoising specifically includes the following steps:

[0011] Step S21: Generalized Anscombe transformation: Performing generalized Anscombe transformation on the wood species image, expressed as: ;in, is the pixel observation value; is the gain coefficient of Poisson noise, , k1 and k2 are coefficients for adjusting noise; is the variance of Gaussian noise, indicating the noise intensity; is the mean of Gaussian noise, indicating the offset of the noise;

[0012] Step S22: Shearlet transform; Use shearlet transform to decompose the wood image and separate high-frequency coefficients containing noise and low-frequency coefficients containing texture; Shearlet transform is expressed as: ; Scaling matrix for shear waves and the shear matrix It is expressed as: ; where SH(·) is the shear wave transfer function; is the low-frequency component of shear wave conversion, is the scaling function; and They are high frequency component and high frequency supplementary component respectively; α is the scale parameter; c is the change control parameter; is the direction parameter; is the wavelet function; is the complementary wavelet function; j is the number of scale decomposition levels; k is the shear amount;

[0013] Step S23: Optimize the threshold function; perform threshold screening on the decomposed high-frequency coefficients, introduce the number of decomposition layers, and perform denoising processing at different decomposition levels adaptively. The dynamic threshold T is expressed as: ; For high frequency coefficients, a threshold function is proposed , expressed as: ; and use the processed high-frequency coefficients and the unprocessed low-frequency coefficients to perform shear wave reconstruction to obtain a denoised image; where, is the basic threshold; N3 is the total number of layers of shear wave decomposition; is the original high-frequency coefficient; sign(·) is the sign function; is the smoothing control coefficient, which determines the intensity of smoothing for coefficients less than the threshold; β is the decay rate parameter, which determines the decay speed of the exponential function and is used to control the amplitude of noise reduction; n is the adjustment coefficient, which is used to control the growth rate; is a correction term; q is a coefficient control term, which controls the processing intensity of coefficients less than the threshold;

[0014] Step S24: inverse conversion: inverse conversion is required after denoising, and the inverse conversion formula is: ;in, is the scale denoised image obtained after inverse transformation; is the denoised image data;

[0015] Step S25: multi-scale reconstruction; fusion of denoising results at different scales to improve denoising accuracy and enhance image details, expressed as: ;in, It is the final denoised image after multi-scale fusion; is the scale weight; N is the total number of scales, i is the scale index; is the image result after denoising at the i-th scale.

[0016] Furthermore, in step S3, the image segmentation specifically includes the following steps:

[0017] Step S31: encoder-decoder structure design; extract high-level semantic features of wood images through convolutional neural networks, including texture, growth rings and microscopic morphological information of the wood surface; the decoder gradually restores the spatial resolution of the feature map extracted by the encoder, and tries to restore the detail information of the wood surface in each step; increase the resolution of the feature map through deconvolution or upsampling operations to generate a feature map of the same size as the input image; introduce a multi-scale feature aggregation module to simultaneously restore features at different scales in each decoding stage, expressed as: ;in, is the feature map of the mth layer; is the upsampling operation; F i (·) is the convolution operation; X is the input;

[0018] Step S32: Hypercolumn fusion; by combining the features of different layers of the encoder, multi-scale information is fused to capture the multi-level features in the wood image, including rough global information and fine local details; specifically: extract feature maps from different levels of the encoder, perform bilinear interpolation on each feature map, and adjust all feature maps to the same resolution; then adjust the number of channels through 1×1 convolution, and fuse all feature maps into an aggregated feature map by element-by-element addition; and add a cross-level feature fusion strategy, specifically, the feature maps of the encoder and decoder are interacted, and the shallow features are complemented with the deep features through cross-layer fusion, ensuring that the fine texture and annual ring information are not lost during the decoding process, which is expressed as: ;in, is the fused feature map; L is the total number of encoder and decoder layers, and l is the layer index; is the feature map extracted by the encoder layer l; is the feature map restored by the decoder layer l; It is an element-by-element addition operation used to fuse the features of the encoder and decoder;

[0019] Step S33: multi-level fusion block; further fuse wood features in channel and spatial dimensions through multi-level fusion blocks; specifically: use global average pooling to compress each channel in the feature map into a scalar to extract global context information; adjust the correlation between channels through the fully connected layer to generate the attention coefficient of each channel; use the Sigmoid function to normalize the attention coefficient to 0 to 1, and apply it to the channels of the original feature map, and introduce a dynamic convolution kernel selection mechanism to perform adaptive weighting between different channels, expressed as: ; The feature map after channel calibration is convolved and fused in the spatial dimension. Through dilated convolution, batch normalization and ReLU activation function, the spatial details in the wood image are gradually restored, which is expressed as: ; Use the second SE block to calibrate the channels of the convolved feature map again to further optimize the attention weights between channels, expressed as: ;in, It is the feature map after channel weighting; is the channel weighting function; DynamicWeight(·) is the dynamic weight used to adjust the feature weights of different channels; Z is the convolution fusion result; BN is batch normalization; It is dilated convolution; is the channel calibration result; It is the channel calibration operation;

[0020] Step S34: segmentation output; use 1×1 convolution to map the fused feature map into a segmentation map of the foreground and background, and the output feature map contains the probability value of each pixel belonging to the foreground;

[0021] Step S35: define a segmentation loss function; ; Where L is the segmentation loss function; C is the total number of categories, and c is the category index; is the class weight; i1 is the pixel index, N1 is the total number of pixels; is the detail weight coefficient, emphasizing the importance of edges and texture areas, , γ is the weight coefficient, is the gradient strength; is the judgment that the i-th pixel belongs to the c-th category, 1 means it belongs, 0 means it does not belong;

[0022] Step S36: classifier training; extracting texture, shape and color features from the segmented image; using a support vector machine classifier to train a model to identify the tree species of the wood.

[0023] Furthermore, in step S4, the wood species recognition is based on the trained vector machine classifier, and the wood species image is collected in real time. After image denoising, image segmentation and feature extraction, the image is input into the trained vector machine classifier, and the wood species classification result output by the vector machine classifier is used as the recognition result.

[0024] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0025] (1) In view of the fact that general wood species recognition methods are unable to adapt to the noise changes in different areas of complex wood species images, there is a problem of loss of edges and details when denoising wood species images, which leads to low final recognition accuracy. This scheme introduces a gain coefficient into the generalized Anscombe transform to more finely control the changes in noise weights of different pixels in wood species images; introduces dynamic threshold screening to compress the noise components below the threshold in wood species images, while avoiding over-smoothing of important information; and obtains the wood species denoising results through multi-scale fusion, which greatly improves the final accuracy of wood species recognition.

[0026] (2) In view of the fact that general wood species recognition methods are not fine enough in restoring information, resulting in insufficient detail representation of wood species images, failure to take into account both global and local information, ignoring subtle texture and color changes of wood, and difficulty in dealing with the problem of imbalanced wood species categories, this scheme introduces a multi-scale feature aggregation module to ensure the delicate restoration of wood species image information. Through a cross-level feature fusion strategy, when processing annual rings and fine lines in wood images, it retains deep semantic information while restoring shallow details. Based on dilated convolution and dynamic convolution kernel selection mechanism, a multi-level fusion block is constructed to help capture annual rings and large texture structures in wood. The segmentation loss function is defined by weight addition to improve the segmentation accuracy of wood species images with imbalanced categories, thereby ensuring the accuracy of the final wood species recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the process of the wood species identification method based on multi-feature fusion provided by the present invention;

[0028] Figure 2 is a schematic flow chart of step S2;

[0029] Figure 3 It is a schematic diagram of the process of step S3.

[0030] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0033] Example 1, see Figure 1 The present invention provides a method for identifying wood species based on multi-feature fusion, the method comprising the following steps:

[0034] Step S1: Image acquisition, collecting historical timber species images;

[0035] Step S2: Image denoising: introducing the gain coefficient into the generalized Anscombe transform to control the noise weight change of different pixels in the wood species image; introducing dynamic threshold screening to compress the noise components below the threshold in the wood species image, thereby achieving image denoising;

[0036] Step S3: Image segmentation, introducing a multi-scale feature aggregation module and a cross-level feature fusion strategy, building a multi-level fusion block based on dilated convolution and dynamic convolution kernel selection mechanism; and defining the segmentation loss function through weight addition, finally achieving image segmentation and completing classifier training;

[0037] Step S4: wood species identification, based on the trained classifier, wood species identification is performed on the wood species images collected in real time.

[0038] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the historical timber species image is a high-resolution image of different types of timber species collected from multiple angles, ensuring that the image covers the full surface features of the timber species, including texture, growth rings and color; the collected historical timber species image is size normalized, and the image is classified and labeled, and the labeled content is the specific category of the timber species.

[0039] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, image denoising specifically includes the following steps:

[0040] Step S21: Generalized Anscombe transformation: Performing generalized Anscombe transformation on the wood species image, expressed as: ;in, is the pixel observation value; is the gain coefficient of Poisson noise, , k1 and k2 are coefficients for adjusting noise; is the variance of Gaussian noise, indicating the noise intensity; is the mean of Gaussian noise, indicating the offset of the noise;

[0041] Step S22: Shearlet transform; Use shearlet transform to decompose the wood image and separate high-frequency coefficients containing noise and low-frequency coefficients containing texture; Shearlet transform is expressed as: ; Scaling matrix for shear waves and the shear matrix It is expressed as: ; where SH(·) is the shear wave transfer function; is the low-frequency component of shear wave conversion, is the scaling function; and They are high frequency component and high frequency supplementary component respectively; α is the scale parameter; c is the change control parameter; is the direction parameter; is the wavelet function; is the complementary wavelet function; j is the number of scale decomposition levels; k is the shear amount;

[0042] Step S23: Optimize the threshold function; perform threshold screening on the decomposed high-frequency coefficients, introduce the number of decomposition layers, and perform denoising processing at different decomposition levels adaptively. The dynamic threshold T is expressed as: ; For high frequency coefficients, a threshold function is proposed , which combines the advantages of hard threshold and soft threshold, solves the discontinuity problem of hard threshold and avoids the over-smoothing problem of soft threshold, and is expressed as: ; and use the processed high-frequency coefficients and the unprocessed low-frequency coefficients to perform shear wave reconstruction to obtain a denoised image; where, is the basic threshold; N3 is the total number of layers of shear wave decomposition; is the original high-frequency coefficient; sign(·) is the sign function; is the smoothing control coefficient, which determines the intensity of smoothing for coefficients less than the threshold; β is the decay rate parameter, which determines the decay speed of the exponential function and is used to control the amplitude of noise reduction; n is the adjustment coefficient, which is used to control the growth rate; is a correction term; q is a coefficient control term, which controls the processing intensity of coefficients less than the threshold;

[0043] Step S24: inverse conversion: After denoising, inverse conversion is required to restore the original numerical distribution of the image. The inverse conversion formula is: ;in, is the scale denoised image obtained after inverse transformation; is the denoised image data;

[0044] Step S25: multi-scale reconstruction; fusion of denoising results at different scales to improve denoising accuracy and enhance image details, expressed as: ;in, It is the final denoised image after multi-scale fusion; is the scale weight; N is the total number of scales, i is the scale index; is the image result after denoising at the i-th scale.

[0045] By performing the above operations, the general wood species recognition method cannot adapt to the noise changes in different areas of complex wood species images, and there is a loss of edges and details when denoising the wood species images, which leads to low final recognition accuracy. This scheme introduces the gain coefficient into the generalized Anscombe transformation to more finely control the noise weight changes of different pixels in the wood species images; introduces dynamic threshold screening to compress the noise components below the threshold in the wood species images, while avoiding excessive smoothing of important information, and obtains the wood species denoising results through multi-scale fusion, which greatly improves the accuracy of the final wood species recognition.

[0046] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, image segmentation specifically includes the following steps:

[0047] Step S31: encoder-decoder structure design; extract high-level semantic features of wood images through convolutional neural networks, including texture, growth rings and microscopic morphological information of the wood surface; the decoder gradually restores the spatial resolution of the feature map extracted by the encoder, and tries to restore the detail information of the wood surface in each step; increase the resolution of the feature map through deconvolution or upsampling operations to generate a feature map of the same size as the input image; introduce a multi-scale feature aggregation module to simultaneously restore features at different scales in each decoding stage, expressed as: ;in, is the feature map of the mth layer; is the upsampling operation; F i (·) is the convolution operation; X is the input;

[0048] Step S32: Hypercolumn fusion; by combining the features of different layers of the encoder, multi-scale information is fused to capture the multi-level features in the wood image, including rough global information and fine local details; specifically: extract feature maps from different levels of the encoder, perform bilinear interpolation on each feature map, and adjust all feature maps to the same resolution; then adjust the number of channels through 1×1 convolution, and fuse all feature maps into an aggregated feature map by element-by-element addition; and add a cross-level feature fusion strategy, specifically, the feature maps of the encoder and decoder are interacted, and the shallow features are complemented with the deep features through cross-layer fusion, ensuring that the fine texture and annual ring information are not lost during the decoding process, which is expressed as: ;in, is the fused feature map; L is the total number of encoder and decoder layers, and l is the layer index; is the feature map extracted by the encoder layer l; is the feature map recovered by the decoder layer l; It is an element-by-element addition operation used to fuse the features of the encoder and decoder;

[0049] Step S33: multi-level fusion block; the multi-level fusion block further fuses wood features in channel and spatial dimensions to solve the semantic difference problem between features at different levels and ensure that the model can effectively integrate multi-scale information; specifically: use global average pooling to compress each channel in the feature map into a scalar to extract global context information; adjust the correlation between channels through the fully connected layer to generate the attention coefficient of each channel; use the Sigmoid function to normalize the attention coefficient to 0 to 1 and apply it to the channel of the original feature map, and introduce a dynamic convolution kernel selection mechanism to perform adaptive weighting between different channels, expressed as: ; The feature map after channel calibration is convolved and fused in the spatial dimension. Through dilated convolution, batch normalization and ReLU activation function, the spatial details in the wood image are gradually restored, which is expressed as: ; Use the second SE block to calibrate the channels of the convolved feature map again to further optimize the attention weights between channels, expressed as: ;in, It is the feature map after channel weighting; is the channel weighting function; DynamicWeight(·) is the dynamic weight used to adjust the feature weights of different channels; Z is the convolution fusion result; BN is batch normalization; It is dilated convolution; is the channel calibration result; It is the channel calibration operation;

[0050] Step S34: segmentation output; use 1×1 convolution to map the fused feature map into a segmentation map of the foreground and background, and the output feature map contains the probability value of each pixel belonging to the foreground;

[0051] Step S35: define a segmentation loss function; ; Where L is the segmentation loss function; C is the total number of categories, and c is the category index; is the class weight; i1 is the pixel index, N1 is the total number of pixels; is the detail weight coefficient, emphasizing the importance of edges and texture areas, , γ is the weight coefficient, is the gradient strength; is the judgment that the i-th pixel belongs to the c-th category, 1 means it belongs, 0 means it does not belong;

[0052] Step S36: classifier training; extracting texture, shape and color features from the segmented image; using a support vector machine classifier to train a model to identify the tree species of the wood.

[0053] By performing the above operations, the general wood species recognition method has the problem that the information recovery is not fine enough, resulting in insufficient detail expression of the wood species image, failure to take into account both global and local information, ignoring the subtle texture and color changes of the wood, and difficulty in dealing with the imbalance of wood species categories. This scheme introduces a multi-scale feature aggregation module to ensure the delicate recovery of wood species image information. Through a cross-level feature fusion strategy, when processing the annual rings and fine lines in the wood image, it retains the deep semantic information while restoring the shallow details; constructs a multi-level fusion block based on the dilated convolution and dynamic convolution kernel selection mechanism to help capture the annual rings and large texture structures in the wood; defines the segmentation loss function through weight addition to improve the segmentation accuracy of wood species images with imbalanced categories; and thus ensures the accuracy of the final wood species recognition.

[0054] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, wood species recognition is based on a trained vector machine classifier. Wood species images are collected in real time. After image denoising, image segmentation and feature extraction, the images are input into the trained vector machine classifier. The wood species classification result output by the vector machine classifier is used as the recognition result.

[0055] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0056] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0057] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A wood species identification method based on multi-feature fusion, characterized by: The method comprises the following steps: Step S1: Image acquisition, collecting historical timber species images; Step S2: image denoising; Step S3: image segmentation; Step S4: wood species identification, based on the trained classifier, wood species identification is performed on the wood species images collected in real time; In step S2, the image denoising comprises the following steps: Step S21: Generalized Anscombe transformation: Performing generalized Anscombe transformation on the wood species image, expressed as: ;in, is the pixel observation value; is the gain coefficient of Poisson noise, , k1 and k2 are coefficients for adjusting noise; is the variance of Gaussian noise, indicating the noise intensity; is the mean of Gaussian noise, indicating the offset of the noise; Step S22: Shearlet transform; Use shearlet transform to decompose the wood image and separate high-frequency coefficients containing noise and low-frequency coefficients containing texture; Shearlet transform is expressed as: ; Scaling matrix for shear waves and the shear matrix It is expressed as: ; where SH(·) is the shear wave transfer function; is the low-frequency component of shear wave conversion, is the scaling function; and They are high frequency component and high frequency supplementary component respectively; α is the scale parameter; c is the change control parameter; is the direction parameter; is the wavelet function; is the complementary wavelet function; j is the number of scale decomposition levels; k is the shear amount; Step S23: Optimize the threshold function; specifically, perform threshold screening on the decomposed high-frequency coefficients, introduce the number of decomposition layers, and perform denoising processing at different decomposition levels adaptively. The dynamic threshold T is expressed as: ; For high frequency coefficients, a threshold function is proposed , expressed as: ; and use the processed high-frequency coefficients and the unprocessed low-frequency coefficients to perform shear wave reconstruction to obtain a denoised image; where, is the basic threshold; N3 is the total number of layers of shear wave decomposition; is the original high-frequency coefficient; sign(·) is the sign function; is the smoothing control coefficient, which determines the intensity of smoothing for coefficients less than the threshold; β is the decay rate parameter, which determines the decay speed of the exponential function and is used to control the amplitude of noise reduction; n is the adjustment coefficient, which is used to control the growth rate; is a correction term; q is a coefficient control term, which controls the processing intensity of coefficients less than the threshold; Step S24: inverse conversion: inverse conversion is required after denoising, and the inverse conversion formula is: ;in, is the scale denoised image obtained after inverse transformation; is the denoised image data; Step S25: multi-scale reconstruction; fusion of denoising results at different scales to improve denoising accuracy and enhance image details, expressed as: ;in, It is the final denoised image after multi-scale fusion; is the scale weight; N is the total number of scales, i is the scale index; is the image result after denoising at the i-th scale; In step S3, image segmentation specifically includes the following steps: Step S31: encoder-decoder structure design; extract high-level semantic features of wood images through convolutional neural networks, including texture, growth rings and microscopic morphological information of the wood surface; the decoder gradually restores the spatial resolution of the feature map extracted by the encoder, and tries to restore the detail information of the wood surface in each step; increase the resolution of the feature map through deconvolution or upsampling operations to generate a feature map of the same size as the input image; introduce a multi-scale feature aggregation module to simultaneously restore features at different scales in each decoding stage, expressed as: ;in, It is the feature map of the m1th layer; is the upsampling operation; F i (·) is the convolution operation; X is the input; Step S32: Hypercolumn fusion; by combining the features of different layers of the encoder, multi-scale information is fused to capture the multi-level features in the wood image, including rough global information and fine local details; specifically: extract feature maps from different levels of the encoder, perform bilinear interpolation on each feature map, and adjust all feature maps to the same resolution; then adjust the number of channels through 1×1 convolution, and fuse all feature maps into an aggregated feature map by element-by-element addition; and add a cross-level feature fusion strategy, specifically, the feature maps of the encoder and decoder are interacted, and the shallow features are complemented with the deep features through cross-layer fusion, ensuring that the fine texture and annual ring information are not lost during the decoding process, which is expressed as: ;in, is the fused feature map; L is the total number of encoder and decoder layers, and l is the layer index; is the feature map extracted by the encoder layer l; is the feature map recovered by the decoder layer l; It is an element-by-element addition operation used to fuse the features of the encoder and decoder; Step S33: multi-level fusion block; further fuse wood features in channel and spatial dimensions through multi-level fusion blocks; specifically: use global average pooling to compress each channel in the feature map into a scalar to extract global context information; adjust the correlation between channels through the fully connected layer to generate the attention coefficient of each channel; use the Sigmoid function to normalize the attention coefficient to 0 to 1, and apply it to the channels of the original feature map, and introduce a dynamic convolution kernel selection mechanism to perform adaptive weighting between different channels, expressed as: ; The feature map after channel calibration is convoluted in the spatial dimension, and the spatial details in the wood image are gradually restored through dilated convolution, batch normalization and ReLU activation function, which is expressed as: ; Use the second SE block to calibrate the channels of the convolved feature map again to further optimize the attention weights between channels, expressed as: ;in, It is the feature map after channel weighting; is the channel weighting function; DynamicWeight(·) is the dynamic weight used to adjust the feature weights of different channels; Z is the convolution fusion result; BN is batch normalization; It is dilated convolution; is the channel calibration result; It is the channel calibration operation; Step S34: segmentation output; use 1×1 convolution to map the fused feature map into a segmentation map of the foreground and background, and the output feature map contains the probability value of each pixel belonging to the foreground; Step S35: define a segmentation loss function; ; Where L is the segmentation loss function; C is the total number of categories, and c is the category index; is the class weight; i1 is the pixel index, N1 is the total number of pixels; is the detail weight coefficient, emphasizing the importance of edges and texture areas, , γ is the weight coefficient, is the gradient strength; is the judgment that the i-th pixel belongs to the c-th category, 1 means it belongs, 0 means it does not belong; Step S36: classifier training; extracting texture, shape and color features from the segmented image; using a support vector machine classifier to train a model to identify the tree species of the wood.

2. The method for identifying wood species based on multi-feature fusion according to claim 1, characterized in that: In step S1, the size of the collected historical wood species images is normalized, and the images are classified and labeled, and the labeled content is the specific category of the wood species.

3. The method for identifying wood species based on multi-feature fusion according to claim 1, characterized in that: In step S4, the wood species recognition is based on the trained vector machine classifier, and the wood species image is collected in real time. After image denoising, image segmentation and feature extraction, the image is input into the trained vector machine classifier, and the wood species classification result output by the vector machine classifier is used as the recognition result.

Images