A grading and sorting system for veneer sheets of wood and its detection method

Abstract

The present invention discloses a grading and sorting system and method for wood rotary cutting thin plates. The grading and sorting system for wood rotary cutting thin plates includes a feeding module, a plate identification and grading module, a plate sorting and packing module, and a conveying module; specifically, it includes a wooden board feeding mechanism, a feeding conveyor belt, a laser measurement sensor, a three-dimensional depth camera, an upper light source, a lower light source, a swing conveyor belt, an output conveyor belt, an output station, a plate alignment mechanism, a plate pushing mechanism, and a bundling mechanism. The plates enter the conveyor belt through the wooden board feeding mechanism. After being identified and graded by the laser measurement sensor and the three-dimensional depth camera, the plates are sorted by the swing electric cylinder and the swing conveyor belt, and finally the grading and sorting of the plates are realized, significantly improving the accuracy and efficiency of plate grading and classification.

Description

Technical Field

[0001] The present invention relates to the field of grading and sorting of tree boards, and particularly to a grading and sorting system for wood veneer sheets and a detection method thereof. Background Art

[0002] In the board processing industry, the grading and classification of boards is a crucial task. Boards of different grades and types are used for different applications, and their prices and demand vary greatly. Therefore, accurately and quickly grading and classifying boards can significantly improve production efficiency, reduce waste, and enhance economic benefits.

[0003] Traditional board grading and classification mostly rely on manual inspection. This method is not only inefficient but also easily affected by individual differences, making it difficult to ensure the consistency and accuracy of the results. In recent years, although some mechanical equipment and methods have been developed, such as using cameras to capture images for analysis, these methods still have certain limitations when dealing with boards with complex defects and cracks. For example, Chinese patent document CN112747788A discloses a "wood board detection device and wood board production line", which uses a dust removal device, a flatness detection device, and a surface defect detection device to achieve the detection of surface defects on both sides of the wood board; a moisture content detection device to achieve the detection of moisture content. These devices and methods can often only detect the surface of relatively thick wood boards and cannot detect defects in wood veneer sheets and accurately measure the thickness of the boards, thus affecting the accuracy of grading and classification of wood veneer sheets. Summary of the Invention

[0004] The object of the present invention is to provide a grading and sorting system for wood veneer sheets and a detection method thereof in view of the deficiencies of the existing technology.

[0005] Technical Solution: The technical solution adopted by the present invention to solve the problem is as follows: A grading and sorting system for wood veneer sheets, the grading and sorting system includes: a feeding module, a board identification and grading module, a board sorting and packing module, and a conveying module; the feeding module includes: a material supporting upward mechanism connected to a board feeding mechanism, a board alignment mechanism installed on the side, and a suction cup feeding mechanism installed above; the board identification and grading module includes: laser measuring devices installed on the upper and lower sides of the feeding conveyor belt, a three-dimensional depth camera and an upper light source installed above the gap between the feeding conveyor belt and the swing conveyor belt, and a lower light source installed below the gap between the feeding conveyor belt and the swing conveyor belt; the board sorting and packing module includes: a number of corresponding discharge conveyor belts and a number of discharge stations, with a material pushing mechanism and a bundling mechanism installed on both sides of the discharge stations respectively; the conveying module includes: a feeding conveyor belt, a swing conveyor belt, and a discharge conveyor belt.

[0006] Preferably, the material supporting and lifting mechanism is connected to the sheet feeding mechanism and periodically and uniformly lifts upward; the sheet aligning mechanism installed on the side periodically aligns and integrates the candidate sheets of the sheet feeding mechanism; the suction cup loading mechanism above uses 4 vacuum rubber suction heads with a diameter of 8 cm to adsorb the candidate sheet materials. After the adsorption is completed, the sheets are moved to the feeding conveyor belt.

[0007] Preferably, the laser measuring devices installed on the upper and lower sides of the feeding conveyor belt detect the position of the sheet by laser beam projection, and while detecting the thickness of the sheet, send a shooting signal to the 3D depth camera.

[0008] Preferably, the swing conveyor belt conveys the sheets to the inlet of the discharging conveyor belt, and the corresponding discharging conveyor belt conveys the sheets to the corresponding discharging stations. According to the stacking situation of the sheets, the material pushing mechanism on one side aligns and pushes the materials, and the material bundling mechanism on the other side bundles and packs the sheets according to a fixed number of sheets.

[0009] Preferably, the feeding conveyor belt, the swing conveyor belt and the discharging conveyor belt in the conveying module are all double-layer conveyor belt structures, and each layer of conveyor belt uses 3 rows of conveyor belts with a width of 2 to 5 cm to run synchronously.

[0010] The present invention also provides a method for grading and sorting veneer sheets of wood, which specifically includes the following steps:

[0011] S1: Perform necessary preprocessing on the depth image obtained by the laser contour depth camera, including but not limited to operations such as noise reduction and normalization;

[0012] S2: Use a generative adversarial network (GAN) to train and generate images with more significant defect features such as live knots, dead knots, cracks, and wrinkles, as well as defect-free positive sample images respectively;

[0013] S3: Perform defect annotation and classification on all real sample images and generated sample images;

[0014] S4: Train a one-stage object detection algorithm model to locate and classify the defects of the sheets, and use the defect recognition results to grade the sheets.

[0015] Preferably, in step S1, the "perform necessary preprocessing on the depth image obtained by the laser contour depth camera, including but not limited to operations such as noise reduction and normalization" is specifically implemented as follows:

[0016] Normalize and reduce the noise of the depth value of the depth map with any bit depth by pixel, and then multiply the depth value by 255, that is, process it into a single-channel grayscale map and mark it as real image data; among them, the noise reduction process uses an improved bilateral filter, and the formula is as follows:

[0017]

[0018] wherein, i and j represent the positions of pixels in the image, and can be expressed as (i x , i y ), (j x , j y ); I(j) represents the pixel value at position j; g(|i - j|) is a spatial Gaussian function, and |i - j| represents the spatial distance between two pixels i and j; h(|I(i) - I(j)|) is an intensity Gaussian function, and |I(i) - I(j)| represents the intensity difference between two pixels i and j, that is, the difference in depth values; for each pixel position i, the influence of the depth values of other pixels j within the ranges of 3×3, 5×5, and 9×9 around it is considered; and the filtering results at the three scales of 3×3, 5×5, and 9×9 are averaged and fused. The average fusion formula is as follows:

[0019]

[0020] wherein, K represents different scales, f(i, k) represents the depth value of the pixel position j at different scales, and the average value is taken after summing the depth values at different scales and the same pixel position.

[0021] Preferably, the specific implementation manner of "using a generative adversarial network (GAN) for training to separately generate images with more significant defect features such as loose knots, dead knots, cracks, and wrinkles, and positive sample images without defects" in step S2 is as follows: performing binary classification on real sample images, which are respectively a first type of defect-free board images and a second type of defective board images, and simultaneously using two GAN networks to read and train the second type of image data to respectively generate a first type of defect-free board images and a second type of defective board images, and marking them as generated image data.

[0022] Preferably, the specific implementation manner of "performing defect annotation and classification on all real sample images and generated sample images" in step S3 is as follows: simultaneously performing defect annotation on real image data and generated image data, including but not limited to defects such as loose knots, dead knots, cracks, and wrinkles; and classifying the real image data and generated image data based on the annotation results by introducing texture and shape features into four categories, namely perfect, excellent, good, and poor.

[0023] Preferably, the specific implementation of "training the one-stage object detection algorithm model to locate and classify the defects of the board and realizing the grading of the board by using the defect recognition results" in step S4 is as follows: using the generated images as the training set, and the real images as the validation set and the test set; training the one-stage object detection deep learning model to recognize the number and type of defects on a single board; combining the results of the four-classification model based on texture and shape features, and using the dynamic weight allocation strategy to realize the grading of the board; the formula of the dynamic weight allocation strategy is as follows:

[0024]

[0025] C l = argmax i (P il )

[0026] In the formula, W i is the probability of the defect recognized by the object detection on the i-th board in the final four-classification; P il represents the probability that the predicted target i belongs to l, represents the probability of the four-classification model based on texture and shape features; finally, the one with the maximum probability is selected as the final classification result C l of the target i.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0028] (1) The present invention realizes the online real-time detection of the thickness and defects of the rotary-cut thin wood board through 3 columns of double-layer conveyor belts, upper and lower light sources, laser measurement sensors and 3D depth cameras, and finally accurately classifies through algorithms, ensuring the accurate classification of the rotary-cut thin wood board.

[0029] (2) The present invention realizes the rapid and efficient implementation of the classification of the classification results of the rotary-cut thin board through the mutual cooperation of the mechanical structures of the swing conveyor belt, the discharge conveyor belt and the discharge station, achieving the one-stop detection and classification of the rotary-cut thin wood board, and greatly improving the production and processing efficiency.

[0030] (3) The present invention combines the detection results of the one-stage object detection deep learning model with the results of the four-classification model based on texture and shape features, avoiding the recognition gap of the algorithm in different batches and different types of boards, improving the robustness of the algorithm, and ensuring the efficient and stable operation of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a side view of the overall structure of the present invention;

[0032] Figure 2 is a top view of the overall structure of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] The present invention will be further illustrated below in conjunction with the accompanying drawings and specific embodiments. These embodiments are implemented on the premise of the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0034] As Figure 1 shown, a grading and sorting system for wood veneer sheets, the grading and sorting system includes: a material supporting and upward moving mechanism 1 is connected to a sheet feeding mechanism 2, a first sheet aligning mechanism 31, a second sheet aligning mechanism 32, and a third sheet aligning mechanism 33 are installed on the side, and a suction cup feeding mechanism 4 is installed above; the sheet identification and grading module includes: a laser measurer 6 installed on the upper and lower sides of a feeding conveyor belt 5, above the gap between the feeding conveyor belt 5 and a swinging conveyor belt 7, a three-dimensional depth camera 8, a first upper light source 91, and a second upper light source 92 are installed, and a lower light source 93 is installed below the gap between the feeding conveyor belt 5 and the swinging conveyor belt 7; the sheet sorting and packing module includes: a first discharging belt 101, a second discharging belt 102, a third discharging belt 103, a fourth discharging belt 104, a first discharging station 111, a second discharging station 112, a third discharging station 113, and a fourth discharging station 114, and a material pushing mechanism 12 and a bundling mechanism 13 are installed on both sides of the first discharging station 111, the second discharging station 112, the third discharging station 113, and the fourth discharging station 114 respectively; the conveying module includes: a feeding conveyor belt 5, a swinging conveyor belt 7, a first discharging belt 101, a second discharging belt 102, a third discharging belt 103, and a fourth discharging belt 104.

[0035] Centered on the end of the feeding conveyor belt 5, the swinging conveyor belt 7 is driven to rotate by a swinging electric cylinder, and the swinging conveyor belt 7 is rotated to the first discharging belt 101, the second discharging belt 102, the third discharging belt 103, and the fourth discharging belt 104 according to the sorting and classification results.

[0036] The feeding conveyor belt 5, the swinging conveyor belt 7, the first discharging belt 101, the second discharging belt 102, the third discharging belt 103, and the fourth discharging belt 104 are all double-layer conveyor belt structures, and each layer of conveyor belt uses 3 columns of conveyor belts with a width of 2 to 5 cm to run synchronously.

[0037] Place the board to be sorted above the board feeding mechanism. The material supporting and lifting mechanism 1 is connected to the board feeding mechanism 2 and periodically and uniformly lifts upward. The first board alignment mechanism 31, the second board alignment mechanism 32, and the third board alignment mechanism 33 installed on the side periodically align and integrate the candidate boards 14 of the board feeding mechanism 2 from three directions. The suction cup feeding mechanism 4 above uses 4 vacuum rubber suction heads with a diameter of 8 cm to adsorb the candidate boards 14. After the adsorption is completed, the top board of the candidate boards 4 is moved to the feeding conveyor belt.

[0038] The laser measuring device 6 installed on the upper and lower sides of the feeding conveyor belt 5 detects the position of the board by laser beam. When it recognizes that the board passes by, it signals the 3D depth camera 8 and the RGB color camera 81 to take pictures simultaneously to complete the detection, classification, and grading of the board thickness and defects.

[0039] According to the classification results, the swing electric cylinder 15 expands and contracts to control the swing conveyor belt 7 to rotate, and convey the board to the first discharge belt 101, the second discharge belt 102, the third discharge belt 103, or the fourth discharge belt 104. Then, at the first discharge station 111, the second discharge station 112, the third discharge station 113, or the fourth discharge station 114, the boards are stacked. For the stacked boards, the material pushing mechanism 12 on one side aligns and pushes the materials, and the material bundling mechanism 13 on the other side bundles and packs the boards according to a fixed number of boards.

[0040] The 3D depth camera 8 and the RGB color camera 81 transmit the captured photos to the image processing terminal through the network cable. The terminal uses the following method to perform grading detection on the boards:

[0041] First, normalize and denoise the depth values of the depth image obtained by the 3D depth camera pixel by pixel. Then multiply the depth values by 255 to process it into a single-channel grayscale image, which is marked as real image data. The denoising process uses an improved bilateral filter, and the formula is as follows:

[0042]

[0043] In the formula, i and j represent the positions of the pixels in the image, which can be expressed as (i x , i y ), (j x , j y); I(j) represents the pixel value at position j; g(|i - j|) is the spatial Gaussian function, where |i - j| represents the spatial distance between two pixels i and j; h(|I(i) - I(j)|) is the intensity Gaussian function, and |I(i) - I(j)| represents the intensity difference between two pixels i and j, that is, the difference in depth values; for each pixel position i, the influence of the depth values of other pixels j within its surrounding ranges of 3×3, 5×5, and 9×9 is considered; and the filtering results in the three scales of 3×3, 5×5, and 9×9 are averaged and fused, and the average fusion formula is as follows:

[0044]

[0045] In the formula, K represents different scales, f(i, k) represents the depth value of pixel position j at different scales, and the average value is taken after summing the depth values at different scales and the same pixel position.

[0046] The processed real sample images are classified into two categories, namely, one category of defect-free board images and two categories of defective board images. Two StyleGANV3 networks are used in parallel to read and train the two-category image data, generating 500 defect-free board images of one category and 500 defective board images of two categories respectively, and they are marked as generated image data.

[0047] Defect annotation is performed on both the real image data and the generated image data, including but not limited to defects such as live knots, dead knots, cracks, and wrinkles; and according to the annotation results, considering both texture and shape features at the same time, the real image data and the generated image data are classified into four categories, namely, perfect, excellent, good, and poor.

[0048] Use the generated images as the training set, and the real images as the validation set and the test set; use the one-stage object detection deep learning model YOLOv7 for training to identify the number and type of defects on a single board; according to the recognition results, use the Softmax function to calculate the probability that the current board belongs to the four-classification, and at the same time combine the results of the four-classification model based on texture and shape features, and use the dynamic weight allocation strategy to achieve the classification of the board; the dynamic weight allocation strategy formula is as follows:

[0049]

[0050] C l = argmax i (P il )

[0051] In the formula, W i is the probability of the defect identified by the object detection of the i-th board in the final four-classification; P il represents the probability that the predicted target i belongs to l, Represents the probability of a four-classification model based on texture and shape features; finally, the one with the highest probability is selected as the final classification result C of target i l 。

[0052] Table 1 shows the result graph of the ablation experiment. The effects of the three methods used in this patent when used separately are accuracy rates of 79.76%, 89.98%, and 85.64% respectively. When the three methods are combined and used simultaneously, first, the depth values of the depth images obtained by the three-dimensional depth camera are normalized and denoised pixel by pixel. Subsequently, two StyleGANV3 networks are used to expand the dataset. After prediction by the one-stage object detection deep learning model YOLOv7, a dynamic weight allocation strategy is used for post-processing, and finally, an accuracy rate of 94.44% is achieved.

[0053] Table 1 Comparison of ablation experiments on the wood veneer grading and sorting method

[0054] Depth image preprocessing GAN network data augmentation Dynamic weight allocation strategy Accuracy √ 79.76％ √ 89.98％ √ 85.64％ √ √ √ 94.44％

Claims

1. A method for grading and sorting veneer sheets of wood, characterized in that: Including the following steps: S1: Preprocess the depth image obtained by the laser profile depth camera. Normalize and denoise the depth values of the depth map with any bit depth on a pixel-by-pixel basis, and then multiply the depth values by 255 to process it into a single-channel grayscale image, which is marked as real image data. Among them, the denoising process uses improved bilateral filtering, and the formula is as follows: wherein, i and j represent the positions of pixels in the image, which can be expressed as (i x , i y ), (j x , j y ); I(j) represents the pixel value at position j; g(|i - j|) is a spatial Gaussian function, and |i - j| represents the spatial distance between two pixels i and j; h(|I(i) - I(j)|) is an intensity Gaussian function, and |I(i) - I(j)| represents the intensity difference between two pixels i and j, that is, the difference in depth values; for each pixel position i, the influence of the depth values of other pixels j within the ranges of 3×3, 5×5, and 9×9 around it is considered; and the filtering results at the three scales of 3×3, 5×5, and 9×9 are averaged and fused, and the average fusion formula is as follows: In the formula, K represents different scales, and f(i,k) represents the depth value of pixel position i at different scales. The depth values at different scales and the same pixel position are summed and then the average value is taken; S2: Use a generative adversarial network (GAN) for training to generate images with more prominent features of live knots, semi-live knots, dead knots, holes, cracks, wrinkles, and knife marks defects respectively, as well as defect-free positive sample images; S3: Perform defect annotation on all real sample images and the generated sample images. According to the annotation results, introduce classification based on texture and shape features for the real image data and the generated image data, which are divided into four categories: perfect, excellent, good, and poor; S4: Train a one-stage object detection algorithm model to locate and classify the defects of the board. Use the defect recognition results to classify the board, including: using the generated images as the training set, and the real images as the validation set and the test set; Train the one-stage object detection deep learning model to identify the number and type of defects on a single board. Combine the results of the four-classification model based on texture and shape features, and use a dynamic weight allocation strategy to classify the board. The formula of the dynamic weight allocation strategy is as follows: C l = argmax i (P il ) Where, W i is the probability of the defect identified by target detection on the i-th plate in the final four-classification; P il represents the probability that the predicted target i belongs to l, represents the probability of the four-classification model based on texture and shape features; the one with the largest probability is finally selected as the final classification result C l of target i.

2. The method for grading and sorting veneer sheets according to claim 1, characterized in that: The step of using the generative adversarial network (GAN) in S2 for training to generate images with more prominent features of live knots, dead knots, cracks, wrinkles and other defect features respectively, as well as defect-free positive sample images includes: performing binary classification on the real sample images, which are divided into two categories: one category is defect-free board images and the other category is defective board images. Then, two GAN networks are used in parallel to read and train the two categories of image data, generating one category of defect-free board images and the other category of defective board images respectively, which are marked as generated image data.

3. A method for grading and sorting veneer sheets of wood according to claim 1, characterized in that: The step of S3 performing defect annotation and classification on all real sample images and the generated sample images includes: performing defect annotation on the real image data and the generated image data simultaneously, including live knots, semi-live knots, dead knots, holes, cracks, wrinkles, and knife marks defects.

Images