WO2019114380A1

WO2019114380A1 - Wood board identification method, machine learning method and device for wood board identification, and electronic device

Info

Publication number: WO2019114380A1
Application number: PCT/CN2018/109106
Authority: WO
Inventors: 丁磊
Original assignee: 北京木业邦科技有限公司
Priority date: 2017-12-14
Filing date: 2018-09-30
Publication date: 2019-06-20
Also published as: CN107944504B; CN107944504A

Abstract

A wood board identification method, a machine learning method and device for wood board identification, and an electronic device. The method comprises: acquiring multiple groups of one-dimensional images of a wood board at a plurality of different preset speeds, wherein each group of one-dimensional images comprises a plurality of one-dimensional images corresponding to a different position of the wood board, and the plurality of one-dimensional images in each group of one-dimensional images correspond to the same preset speed (S101); splicing each group of one-dimensional images in the obtained multiple groups of one-dimensional images, so as to obtain a plurality of two-dimensional images at the different preset speeds (S102); and training wood board identification models respectively using the types of wood boards, the plurality of different preset speeds and the plurality of two-dimensional images as multiple groups of training data (S103), wherein each group of training data in the multiple groups of training data comprises the type of a wood board, one of the two-dimensional images and a corresponding preset speed, and an identification result of a wood board identification model comprises the type of the wood board and the moving speed.

Description

Machine learning method, device and electronic device for wood board recognition and wood board recognition

Cross-reference to related applications

The present application claims the priority of the Chinese Patent Application Serial No.

Technical field

The present disclosure relates to the field of wood automated processing technologies, and in particular, to a machine learning method, apparatus, electronic device, and computer readable storage medium for wood board identification and board recognition.

Background technique

In the field of wood processing, board sorting is an important part. Whether it is a semi-finished product or a finished product after molding, coloring, drying, etc., it is necessary to classify according to different wood characteristics combined with quality standards. In the conventional method, the sorting of the board is done manually. Trained workers, through observation, judge the color, texture, defects, etc. of each piece of wood, and combine the experience to classify a piece of wood into different categories. The wood boards in each category have closer characteristics, achieving a higher product appearance and consistency of quality.

However, the use of manual sorting methods requires a lot of human resources, and since each batch of wood material and coloring process may be different, each product classification standard may also change, so it is necessary to constantly Conduct training and training. At the same time, with the increase of working hours, the manpower method will also have a phenomenon of decreasing accuracy and slowing down efficiency.

The use of machines for wood sorting is becoming an emerging trend in the current industry, and many steps in the wood processing can be solved by machine methods. However, most of these techniques use a fixed method to extract the features of wood or wood. The sorting parameters and methods are fixed and must be designed and adjusted to operate effectively. Therefore, there are certain defects in standardization and applicability.

With the recent advances in machine learning, the use of machine learning for automated board processing has become increasingly popular. A wood classification method using machine learning has also appeared in the related art, and images of pre-specified sample wood are collected by various imaging methods, and then a model is trained using a machine learning method, and automatic classification detection is realized by the model.

However, the inventors invented in the process of implementing the present invention that the existing machine learning method only partially solved the problem of standardization and failed to solve the most critical adaptation problem in wood classification. Specifically, the current standard for wood board classification is usually customized by the manufacturer, that is, the current board classification is essentially non-standardized; the machine learning method in the related art adopts a pre-training method, and only a specific model can be trained for a specific manufacturer. Obviously, it cannot be used in multiple vendors. In addition, the wood is produced in batches. Each batch of products is highly correlated with the original wood quality and painting process of the batch. There may be huge differences between batches; however, the training model in the related art is completely Depending on the sample lot, dynamic tuning is not possible for different batches. Finally, the change of natural light will have a greater impact on visual recognition. The related technology does not consider the standardization under ambient light changes and cannot adapt to the detection of different environments.

Therefore, there is currently no adaptive detection method, and the related technology cannot meet the requirements of rapid deployment, nor can it operate robustly in a variable operating environment.

Summary of the invention

Embodiments of the present disclosure provide a machine learning method, apparatus, and computer readable storage medium for wood board recognition.

In a first aspect, a machine learning method for wood board recognition is provided in an embodiment of the present disclosure.

Obtaining a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds; wherein each set of one-dimensional images includes a plurality of one-dimensional images corresponding to different positions of the wooden board, and the plurality of one-dimensional images in each set of one-dimensional images correspond to The same predetermined speed;

Splicing each set of one-dimensional images in the acquired plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

The board identification type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as the plurality of sets of training data to respectively train the board recognition model; each group of the training data includes the category of the board, and the plurality of two-dimensional images. A two-dimensional image and corresponding predetermined speed; the recognition result of the wood board recognition model includes the category of the board and the moving speed.

Optionally, acquiring a plurality of sets of one-dimensional images of the wood board at a plurality of different predetermined speeds, including:

In the case where the board has a relative speed with the linear camera, and the combination of the relative speed and the sampling frame rate of the linear camera corresponds to a plurality of different predetermined speeds, a plurality of sets of one-dimensional images of the board are acquired.

Obtaining a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds under different illumination conditions;

The wood board identification model is trained by using the category of the board, a plurality of different predetermined speeds, and a plurality of two-dimensional images as the plurality of sets of training data, including:

The board identification type, a plurality of different predetermined speeds, different illumination conditions, and a plurality of two-dimensional images are used as a plurality of sets of training data to respectively train the wood board recognition model; each group of training data includes a category of wood boards, and multiple A two-dimensional image in a two-dimensional image and its corresponding predetermined velocity and illumination conditions.

Optionally, each of the acquired one-dimensional images in the plurality of sets of one-dimensional images is separately spliced to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds, including:

Dividing each set of one-dimensional images of the acquired plurality of sets of one-dimensional images into at least two groups according to an image acquisition time sequence and/or an image order;

The one-dimensional images of at least two groups are separately spliced to form a plurality of two-dimensional images at different predetermined speeds.

Optionally, the lighting condition includes one or more of a strength of the light source outside the board, a direction of illumination of the external source light, a shooting angle of the image acquiring unit that acquires the one-dimensional image, and an aperture size of the image acquiring unit; the predetermined speed is The relative movement speed between the image acquisition unit and the board.

Optionally, the method further includes:

A plurality of one-dimensional images of white reference objects are also acquired while acquiring a plurality of sets of one-dimensional images of the wood board.

Optionally, the board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as the plurality of sets of training data to respectively train the board recognition model, including:

Combining each set of one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images;

A plurality of two-dimensional images are respectively labeled to obtain boundary information of the wooden board.

Optionally, the board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as the plurality of sets of training data to respectively train the board recognition model, and further includes:

The board boundary recognition model is trained according to the boundary information of the plurality of two-dimensional images and the wooden board, and the recognition result of the board boundary recognition model includes the boundary information of the board.

In a second aspect, an embodiment of the present disclosure provides a board identification method, including:

Obtaining multiple one-dimensional images of the board;

Splicing the acquired one-dimensional images to obtain a two-dimensional image to be identified;

The recognition is based on the two-dimensional image and the trained wood board recognition model, and the type of the board and the moving speed are obtained.

Optionally, the obtained one-dimensional images are spliced to obtain a two-dimensional image to be identified, including:

Dividing the acquired plurality of one-dimensional images into at least two sets of one-dimensional images according to an image acquisition time sequence and/or an image order;

At least two sets of one-dimensional images are separately spliced to form at least two two-dimensional images to be identified.

Optionally, identifying according to the two-dimensional image and the trained wood board recognition model, obtaining the category of the board and the moving speed, including:

Inputting at least two two-dimensional images into the wood board recognition model respectively, obtaining two types of confidence estimates of the category of the board and the moving speed;

The group with the highest confidence is selected from the two sets of confidence estimates as the final recognition result.

Optionally, the method further includes:

The kick timing of the board is obtained according to the moving speed of the board.

Optionally, the two-dimensional image and the trained wood board recognition model are identified to obtain the category of the board and the moving speed, including:

Obtaining boundary information of the board based on the two-dimensional image and the trained boundary recognition model;

According to the two-dimensional image, the boundary information, and the wood board recognition model, the category of the board and the moving speed are obtained.

A plurality of two-dimensional images obtained under different illumination conditions are respectively input to the wood board recognition model to obtain a plurality of sets of confidence estimates of the category of the board and the moving speed;

The group with the highest confidence is selected from the multi-group confidence estimates as the final recognition result.

Optionally, the method further includes:

A plurality of one-dimensional images of white reference objects are also acquired while acquiring a plurality of one-dimensional images of the wood board.

In a third aspect, a machine learning apparatus for wood board identification is provided, comprising:

a first obtaining module configured to acquire a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds; wherein each set of one-dimensional images includes a plurality of one-dimensional images at different positions of the corresponding wooden board, and each set of one-dimensional images Multiple one-dimensional images in the image correspond to the same predetermined speed;

The first splicing module is configured to splicing each set of one-dimensional images in the acquired plurality of one-dimensional images to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

The training module is configured to train the wood board recognition model as a plurality of sets of training data by using a category of the board, a plurality of different predetermined speeds, and a plurality of two-dimensional images; each group of the training data includes a category of the board, a two-dimensional image of the plurality of two-dimensional images and a corresponding predetermined speed; the recognition result of the wood board recognition model includes a category of the wood board and a moving speed.

Optionally, the first obtaining module includes:

The first acquisition sub-module is configured to acquire a plurality of sets of one-dimensional images of the wood board in a case where the board has a relative speed with the linear camera and the combination of the relative speed and the sampling frame rate of the linear camera corresponds to a plurality of different predetermined speeds.

Optionally, the first obtaining module includes:

a second obtaining sub-module configured to acquire a plurality of sets of one-dimensional images of the wood board at a plurality of different predetermined speeds under different lighting conditions;

Training modules, including:

The first training sub-module is configured to train the wood board recognition model as a plurality of sets of training data by using a category of the board, a plurality of different predetermined speeds, different illumination conditions, and a plurality of two-dimensional images; each group of the plurality of sets of training data The training data includes the category of the board, a two-dimensional image of the plurality of two-dimensional images, and their corresponding predetermined speeds and lighting conditions.

Optionally, the first splicing module includes:

a first grouping sub-module configured to divide each set of one-dimensional images in the acquired plurality of sets of one-dimensional images into at least two groups according to an image acquisition time sequence and/or an image order;

The first splicing sub-module is configured to splicing the one-dimensional images of the at least two groups separately to form a plurality of two-dimensional images at different predetermined speeds.

Optionally, the device further includes:

The second obtaining module is configured to acquire a plurality of one-dimensional images of the white reference object while acquiring the plurality of sets of one-dimensional images of the wooden board.

Optionally, the first splicing module includes:

a second splicing sub-module configured to splicing each set of one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images;

The labeling sub-module is configured to label a plurality of two-dimensional images separately to obtain boundary information of the board.

Optionally, the training module further includes:

The second training sub-module is configured to train the board boundary recognition model according to the boundary information of the plurality of two-dimensional images and the wooden board, and the recognition result of the board boundary recognition model includes the boundary information of the board.

The functions can be implemented in hardware or in hardware by executing the corresponding software. The hardware or software includes one or more modules corresponding to the functions described above.

In one possible design, the structure of the board-recognized machine learning device includes a memory and a processor for storing one or more machine learning devices that support wood board recognition to perform the machine learning method of board identification in the first aspect above. Computer instructions, the processor being configured to execute computer instructions stored in the memory. The board-aware machine learning device may also include a communication interface with which the machine learning device for wood board identification communicates with other devices or communication networks.

In a fourth aspect, an embodiment of the present disclosure provides a board identifying device, including:

a third obtaining module configured to acquire a plurality of one-dimensional images of the wood board;

a second splicing module configured to splicing the acquired one-dimensional images to obtain a two-dimensional image to be recognized;

The identification module, the splicing module identifies the two-dimensional image and the trained wood board recognition model, and obtains the category of the board and the moving speed.

Optionally, the second splicing module includes:

a second grouping sub-module configured to divide the acquired plurality of one-dimensional images into at least two sets of one-dimensional images according to an image acquisition time sequence and/or an image order;

The third splicing sub-module is configured to splicing at least two sets of one-dimensional images to form at least two two-dimensional images to be identified.

Optionally, the identification module includes:

a first identification sub-module configured to input at least two two-dimensional images into the wood board recognition model respectively, to obtain two types of confidence estimates of the category of the board and the moving speed;

The first selection sub-module is configured to select the group with the highest confidence from the two sets of confidence estimates as the final recognition result.

Optionally, the method further includes:

The kick leg module is configured to obtain the kick timing of the board based on the moving speed of the board.

Optionally, the identification module includes:

a second identification sub-module configured to obtain boundary information of the wooden board according to the two-dimensional image and the trained boundary recognition model;

The third identification sub-module is configured to obtain the category of the wooden board and the moving speed according to the two-dimensional image, the boundary information, and the wood board recognition model.

Optionally, the identification module includes:

a fourth identification sub-module configured to input a plurality of two-dimensional images obtained under a plurality of different illumination conditions into the wood board recognition model respectively, to obtain a plurality of sets of confidence estimates of the category of the board and the moving speed;

The second selection sub-module is configured to select the group with the highest confidence from the plurality of sets of confidence estimates as the final recognition result.

Optionally, the method further includes:

The fourth obtaining module is configured to acquire a plurality of one-dimensional images of the white reference object while acquiring the plurality of one-dimensional images of the wood board.

In a possible design, the structure of the board identifying device includes a memory and a processor for storing one or more computer instructions for supporting the board identifying device to perform the board identifying method in the first aspect, the processor being configured to use Computer instructions stored in the execution memory. The board identification device may also include a communication interface for the board identification device to communicate with other devices or communication networks.

In a fifth aspect, an embodiment of the present disclosure provides an electronic device, including a memory and a processor; wherein the memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the first Aspect or method step of the second aspect.

In a sixth aspect, an embodiment of the present disclosure provides a computer readable storage medium for storing computer instructions for a wood board identification machine learning device or a wood board identification device, including a machine for performing wood board recognition in the above first aspect The computer method involved in the learning method or the second aspect of the board identification method.

The technical solutions provided by the embodiments of the present disclosure may include the following beneficial effects:

In the embodiment of the present disclosure, a plurality of sets of one-dimensional images of wood boards at different speeds are acquired by a linear camera, and a plurality of sets of one-dimensional images are further spliced to form a plurality of two-dimensional images corresponding to different speeds, thereby utilizing speed, real categories of wood boards, and corresponding The two-dimensional image respectively trains the wood board recognition model, so that the wood board recognition model after the training can automatically recognize the type and speed of the board. The embodiment of the present disclosure collects a one-dimensional image of a wooden board at different speeds, and uses the same wooden board image samples at different speeds, the type of the wooden board, and the speed to train the machine learning model, so that the wood board recognition model obtained by the training can be accurately The image of the wood board collected at different speeds is recognized, and the relative movement speed of the board is also recognized. The wood board recognition model obtained by the present disclosure can accurately classify the board regardless of the environment and any speed, and perform the sorting operation at the exact kick time. Moreover, there is no need for a separate device, such as a photoelectric sensor, to obtain the time of arrival of the board, but the classification time can be determined by classification learning.

The above general description and the following detailed description are intended to be illustrative and not restrictive.

DRAWINGS

Other features, objects, and advantages of the present invention will become more apparent from the aspects of the appended claims. In the drawing:

1 shows a flow chart of a machine learning method for wood board recognition in accordance with an embodiment of the present disclosure;

A schematic diagram of a convolutional neural network in accordance with an embodiment of the present disclosure is shown in FIG. 2;

3 is a schematic view showing the overall structure of a wood sorting system according to an embodiment of the present disclosure;

Figure 4 shows a flow chart of step S102 according to the embodiment shown in Figure 1;

Figure 5 shows a flow chart of step S103 according to the embodiment shown in Figure 1;

FIG. 6 illustrates an example of performing boundary labeling on an image sample according to an embodiment of the present disclosure;

FIG. 7 illustrates a flowchart of a wood board identification method according to an embodiment of the present disclosure;

Figure 8 shows a flow chart of step S702 according to the embodiment shown in Figure 7;

9 is a block diagram showing the structure of a machine learning device for wood board recognition according to an embodiment of the present disclosure;

FIG. 10 is a block diagram showing the structure of a wood board identifying apparatus according to an embodiment of the present disclosure; FIG.

11 is a block diagram showing the structure of an electronic device suitable for implementing a machine learning method for wood board recognition according to an embodiment of the present disclosure.

Detailed ways

In prior art machine learning applications, the camera is required to use low precision cameras as much as possible to reduce costs. This is because some classification algorithms do not require very high precision image data to achieve accurate classification. However, in the field of wood processing, due to the very slight differences in the texture and color of the board, the machine learning method has higher requirements on the resolution of the camera. Conventional cameras use rectangular sensors, which have the advantage of being able to obtain two-dimensional image data that is easy to process. However, this also limits the resolution of the sample. Linear cameras have higher one-dimensional resolution and variable sampling periods, so they have better application potential in wood processing.

FIG. 1 illustrates a flow chart of a machine learning method for wood board recognition in accordance with an embodiment of the present disclosure. As shown in FIG. 1, the machine learning method for wood board recognition includes the following steps S101-S103:

In step S101, a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds are acquired; wherein each set of one-dimensional images includes a plurality of one-dimensional images at different positions of the corresponding wooden boards, and each of the one-dimensional images in each group One-dimensional images correspond to the same predetermined speed;

In step S102, each set of one-dimensional images in the acquired plurality of sets of one-dimensional images are separately spliced to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

In step S103, the board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as the plurality of sets of training data to respectively train the board recognition model; each group of the training data includes the category of the board, and more A two-dimensional image of the two-dimensional image and the corresponding predetermined speed; the recognition result of the wood-board recognition model includes the category of the wood board and the moving speed.

In the embodiment, when the machine learning is performed by using the image of the wood board sample, the collected wood board image is a high-precision one-dimensional image, and images of different moving speeds are collected for the same wood board sample, so that the training sample is more abundant, and Can adapt to the recognition of images acquired under various conditions. The different moving speeds here refer to the relative speed between the board and the image acquisition device, ie the linear camera.

In this embodiment, the one-dimensional image can be acquired by a linear camera. Since the linear camera can only acquire one one-dimensional image at each time point, the one-dimensional image cannot be directly used for subsequent learning or classification. Therefore, a plurality of one-dimensional data can be spliced into one two-dimensional image, and the two-dimensional image can contain image information of a part or the whole board.

In this embodiment, the predetermined speed may be a relative speed between a predefined board and a linear camera. When collecting a one-dimensional image of a wooden board, when the relative speed between the wooden board and the linear camera is a predetermined speed, a one-dimensional image of the wooden board is collected, and one one-dimensional image is collected at each time point in sequence, and more is collected at the same predetermined speed. One-dimensional images are a set of one-dimensional images. After collecting a plurality of one-dimensional images of part or the whole board, the relative speed between the board and the linear camera can be changed, and the acquisition is performed again, and finally a plurality of sets of one-dimensional images at different predetermined speeds are obtained. By splicing a set of one-dimensional images acquired at the same predetermined speed, a two-dimensional image corresponding to the predetermined speed is obtained, and finally a plurality of two-dimensional images corresponding to different predetermined speeds can be obtained.

In the embodiment, when training the wood board recognition model, the category of the board, the predetermined speed, and the two-dimensional image corresponding to the predetermined speed are used as a set of training data, and the same board corresponds to multiple sets of training data, and each group of training data The two-dimensional image is used as an input to the machine learning model, and the corresponding predetermined speed and board type are used as outputs to train various parameters of the machine learning model. After training through multiple sets of training data corresponding to multiple planks, the result of the machine learning model is converged, and finally the trained wood board recognition model is obtained.

The machine learning model may include, but is not limited to, one or a combination of a convolutional neural network, a feedback neural network, a deep learning network, a decision forest, a Bayesian network, a support vector machine.

The neural network is taken as an example to introduce the principle and process of model training in detail.

The neural network may be an automatic classification model, a regression model or a decision model, and the neural network may be one or a combination of a convolutional neural network and a deep neural network. The neural network may comprise a neural network comprising a plurality of layers, each layer comprising a plurality of nodes, and a trainable weight (i.e., parameters of the aforementioned machine learning model) between the two adjacent nodes.

A schematic of a convolutional neural network is shown in Figure 2, which includes multiple convolutional and downsampled layers and a fully connected layer. The convolutional layer is the core module of the convolutional neural network. By convolving with a filter, multiple nodes of the previous layer are connected to the nodes of the next layer. In general, each node of the convolutional layer is only connected to a part of the nodes of the previous layer. Through the training process, the filter using the initial value can continuously change its own weight according to the training data, and then generate the final filter value. The downsampling layer can use a max-pooling method to reduce a set of nodes into one node, using a method of nonlinearity taking the maximum value. After passing through multiple convolutional and downsampled layers, a fully connected layer is ultimately used to generate the output of the classification. The fully connected layer connects all nodes of the previous layer to all nodes of the latter layer, which is related to a traditional neural The network is similar.

For example, during training, a training algorithm, such as a gradient descent algorithm, can be used to change the filter weight values in the neural network, thereby minimizing the difference in classification between the output and the sample data. As the amount of training data used continues to increase, the changing network node values are constantly changing and improving, and the classification ability of the neural network is also improved. The training process can be done locally or in the cloud. In the case of training in the cloud, the determined classification of the board and the relationship between the 2D image and the custom classification, as well as the 2D image, etc., can be uploaded to the cloud. The cloud server uses the obtained custom classification, the relationship between the two-dimensional image and the custom classification, and the two-dimensional image to train the neural network, and deploys the trained wood board recognition model to the local.

In this embodiment, the category of the board can be predetermined. First determine the board samples, then customize the board samples, for example, classify the 1-3 board samples into Class A categories, class 4-8 boards into Class B categories, and Class 9-10 boards into Class C categories. . Because it is a custom classification, it can be customized according to the specific conditions of the board factory and the actual classification. For example, the boards of No. 1, 3, and 5 are classified into Class A categories, and the remaining board samples are classified into Class B categories. In the factory, custom classification is performed according to the actual situation of the factory. This kind of custom classification is more suitable for the actual situation of different board factories and the classification requirements, and the classification is more flexible and convenient. The specific implementation of the classification is done manually by experience. How many categories are set and which sample belongs to which category is also implemented manually. Manual sorting can be done based on different features of the board, such as color, texture, defects, and the like.

In an optional implementation manner of this embodiment, the step S103 is to obtain a plurality of sets of one-dimensional images of the board at a plurality of different predetermined speeds, and further includes:

In this embodiment, when sampling a one-dimensional image of a wooden board at a predetermined speed, a plurality of methods can be employed. One way is to move the board to the linear camera with the linear camera fixed. For example, the movement of the board is achieved by a conveyor belt, and a linear camera is mounted above the conveyor belt to capture images. 3 is a block diagram showing the structure of a wood board sorting system for sampling wooden board images in an embodiment of the present disclosure. The structural details of Figure 3 will be described in detail in the section on Woodboard Classification and Identification. The wood board sample enters the finishing area B1, in which the wood board samples are finished, guided, and then sequentially entered into the determination area B2 (image sampling area). In the B2 area, one or more linear cameras on the pipeline capture the one-dimensional image of the wood sample at a predetermined speed with a very rapid speed. The conveyor moves at a different predetermined speed each time: V1 represents a predetermined speed of 0, V2. V3....Vn represents the speed before the predetermined speed is from 0 to the line belt of the B2 area, and V(n+1) represents the B2 line speed in order to obtain a one-dimensional image of the board at different predetermined speeds.

Alternatively, the linear camera moves toward the plank sample with the plank sample fixed. Similarly, one or more linear cameras scan the wood sample at a predetermined speed and capture the image of the wood sample at a very fast rate: V1 represents speed 0, V2, V3....Vn represents 0 to B2, respectively. The velocity before the regional pipeline belt is synchronized, V(n+1) represents the B2 pipeline speed, in order to obtain a one-dimensional image of the plank samples at different speeds. This method is suitable for cases where the wooden board sample is large and inconvenient to move.

In other ways, since the frame rate of the linear camera can be adjusted, the moving speed can also be simulated by adjusting the frame rate of the linear camera. For example, when the board moves and the linear camera is fixed, the first frame rate is used to collect the one-dimensional image in the first half of the board sample, and the second frame rate is used to collect the one-dimensional image in the second half of the board sample, so that two different differences of the same board can be obtained. An image sample of a predetermined speed. For another example, the board is fixed, and when the linear camera is moved, the image samples at various predetermined speeds can also be obtained by adjusting the frame rate of the linear camera. For example, a linear camera moves the board several times at the same speed V, but scans the sample using different frame rates, for example, f0, f2, f3....fn, and the obtained board image and the board obtained at different moving speeds. The image is the same.

Of course, in other optional embodiments, both the wood sample and the linear camera may be moved, and the moving direction may be relative or the same direction, as long as the relative motion speed between the two is not zero (ie, At the same time, the linear camera uses a certain frame rate to scan the wood sample, and the aforementioned predetermined speed is formed by the combination of the relative motion speed and the sampling frame rate. Among them, the moving speed of the wood board sample, the moving speed of the linear camera, the sampling frame rate, and the predetermined speed are all variable. For example, when the plank sample is relatively moved at the first speed v1 and the linear camera at the second speed v2, the relative motion speed of the board is v1+v2, and the one-dimensional image obtained by the linear camera using the standard frame rate scan is predetermined. For a one-dimensional image with a speed of v1+v2, a one-dimensional image obtained by scanning at a standard frame rate of 2 times is a one-dimensional image at a predetermined speed of (v1+v2)/2, and is scanned using a 1/2 standard frame rate. The one-dimensional image is a one-dimensional image at a predetermined speed of 2 (v1+v2). That is, in the embodiment of the present application, the predetermined speed can be obtained by adjusting one or more of the moving speed of the wood board sample, the moving speed of the linear camera, and the sampling frame rate.

In an optional implementation manner of this embodiment, the step S101 is to obtain a plurality of sets of one-dimensional images of the board at a plurality of different predetermined speeds, including:

In this optional implementation, image acquisition under different illumination conditions is achieved by transforming the aperture of the linear camera during image acquisition; it is also possible to set the light source around the linear camera to achieve different brightness by adjusting the brightness of the light source or the illumination direction of the light source. The lighting conditions. For example, as shown in FIG. 3, one or more light sources are added to the B2 area production line, for example, the light source may be a flat type LED lamp: L1, L2.....Ln, the LED lamp can provide relatively uniform illumination, and The brightness of the light source can be sequentially increased or decreased by the control method to obtain samples of the product under different light conditions; the aperture size and the brightness or illumination direction of the light source can be simultaneously changed to achieve different illumination conditions. In one mode, the illumination condition of the LED lamp is used to raise the basic brightness of the image acquisition to a satisfactory level, and at the same time, by changing the size of the aperture to obtain the illumination level of the upper and lower floating of the basic brightness. The combination of the two enables multiple illumination conditions to be achieved within a satisfactory range. Since the linear camera only collects one-dimensional images in a unit time, the entire board is multi-collected and finally spliced into a two-dimensional image. Therefore, different illumination intensities can be used at different times by a method synchronized with the illumination device. For example, at t1 sampling time, s1 illumination intensity is used; at t2 sampling time, s2 illumination intensity is used; t3 sampling time is used, s3 illumination intensity is used, and so on, and multiple one-dimensional images are obtained. The images of the odd moments are assembled into a first image sample, the even time instants are assembled into a second image sample, and after the wood board samples are scanned, two image samples under two different illumination conditions are obtained. Alternatively, the s1 light intensity can be used during the first half of the scan area of the board sample, and the s2 light intensity can be used during the second half of the scan area of the board sample, or image samples at two light intensities can be obtained. This method of changing the illumination intensity can be used in conjunction with the method of changing the aperture to obtain image samples with more illumination intensity.

In an optional implementation manner of the embodiment, as shown in FIG. 4, in step S102, each set of one-dimensional images in the acquired plurality of one-dimensional images is separately spliced to obtain a plurality of different predetermined speeds. The step of multiple two-dimensional images further includes the following steps:

In step S401, each set of one-dimensional images in the acquired plurality of sets of one-dimensional images are respectively divided into at least two groups according to an image acquisition time sequence and/or an image order;

In step S402, at least two groups of one-dimensional images are separately spliced to form a plurality of two-dimensional images at different predetermined speeds.

In this optional implementation, multiple one-dimensional images in the same group can be divided into at least two groups by chronological order of image acquisition; for example, the first half of the same group of one-dimensional images are divided into one group, and the second half is one-dimensional. The images are divided into a group; the plurality of one-dimensional images in the same group can also be divided into at least two groups by image order; for example, one-dimensional images in the same group with an odd order are divided into one group, and the order is evenly divided into one group. Of course, the same group of one-dimensional images can be divided into multiple groups according to other methods, which can be set according to actual conditions, and details are not described herein again.

For example, the first and second parts of the obtained one-dimensional image are divided into two parts according to the chronological order, and finally the one-dimensional images of the front and the back are respectively spliced to obtain two one-dimensional images. At different predetermined speeds, 2N two-dimensional images will be obtained, and the number of groups in N and multiple sets of one-dimensional images is the same. As before, when the first half of the one-dimensional image and the second half of the one-dimensional image are sampled, different illumination conditions can be used, so that the two two-dimensional images obtained have the same predetermined speed, but the illumination conditions are different.

For another example, the acquired set of one-dimensional images may be divided into two parts according to the image order, that is, the one-dimensional images of the odd-numbered and even-numbered ones in the set of one-dimensional images acquired at the same predetermined speed are respectively divided into two groups. The one-dimensional images included in each group are spliced into a two-dimensional image, and two two-dimensional images can be obtained at the same predetermined speed. At different speeds, a 2N two-dimensional image can be obtained. As before, when sampling and acquiring a one-dimensional image at an odd time and sampling a one-dimensional image at an even time, different illumination conditions can be used, so that the two two-dimensional images obtained have the same predetermined speed, but the illumination conditions are different.

In an optional implementation manner disclosed in this embodiment, the method further includes:

In this optional implementation, a reference image can be set during image acquisition. For example, in the image acquisition area, a white reference object is provided to ensure that the image of the wood sample is captured simultaneously with the image of the white reference object. A white reference object can be used to provide a reference for white balance, brightness, or other image parameters.

In summary, before the machine learning model, what is obtained is a set of training data, each set of training data includes at least one board category (ie, a customized product category) and a speed, and may also include a lighting condition and/or A camera angle label. For example, the following is a set of training data that is ultimately used for subsequent learning steps:

Sample 1

[Category: A, Speed: V2, Light intensity: L3, Camera angle, A5]

Sample 2

[Category: A, Speed: V3, Light intensity: L3, Camera angle, A5]

Sample 3

[Category: B, Speed: V0, Light intensity: L3, Camera angle, A5]

Sample 4

[Category: A, Speed: V2, Light intensity: L3, Camera angle, A5]

Among them, due to the use of a linear camera, a Sample can be a combination of multiple one-dimensional image data. The speed information can be determined by the true relative movement speed and the frame rate. Multiple images Sample can be obtained from the same board.

In an optional implementation manner of this embodiment, as shown in FIG. 5, step S103 is to train the wood board recognition model as a plurality of sets of training data, that is, a category of the board, a plurality of different predetermined speeds, and a plurality of two-dimensional images. The steps further include:

In step S501, each set of one-dimensional images in the plurality of sets of one-dimensional images are separately spliced to obtain a plurality of two-dimensional images;

In step S502, a plurality of two-dimensional images are respectively labeled to obtain boundary information of the board.

In this optional implementation, each image sample is a two-dimensional image sample, and the image sample itself contains a boundary information, that is, how the one-dimensional image data obtained by the linear camera is uninterrupted is cut and spliced into an independent A sample of a two-dimensional image. An annotation uses the image itself as a boundary, that is, without any additional annotations, using only one independent two-dimensional image that contains both a woodboard image and a useless background image as a sample. Another type of annotation uses an additional boundary label to independently mark the boundaries of the board image to distinguish the board image from the background image. Figure 6 shows an example of boundary labeling of image samples, where the start and end boundaries are used to determine a separate board image and the side borders are used to determine the boundaries between the board image and the background image. The side boundaries can be distinguished from the start and end. Since the linear camera can only output the stitched image without interruption, cutting and labeling need to be done manually. It is to be understood that the present disclosure is not limited to the one of the boundary labeling methods shown in FIG. 6. In an optional implementation manner of the embodiment, the step S103, that is, the step of training the board recognition model by using the category of the board, the plurality of different predetermined speeds, and the plurality of two-dimensional images as the plurality of sets of training data, further includes:

In this optional implementation, the linear camera can continuously acquire one-dimensional images of a plurality of wood boards moving or arranged in a pipeline, and splicing to obtain a continuous two-dimensional image. In this case, a cutting model, that is, a board boundary recognition model, can be trained to cut a continuous two-dimensional image, that is, to identify boundary information in the two-dimensional image, and to divide the boundary information into a plurality of two-dimensional images. Get a 2D image sample each containing only one board image. The cutting model can use a separate neural network or share a neural network with the wood board recognition model. In order to facilitate the distinction between the board and the board identification network, the function of the network is described only in terms of functionality. In actual implementation, there may not be a separate entity or output value. The cutting model can be trained by the above-mentioned image data marked with a boundary to obtain a decision strategy. After receiving the uninterrupted two-dimensional image, a decision of the starting boundary can be generated at time t1, and an end is generated at time t2. The judgment of the border. Further, the image data at times t1 and t2 is composed into an image containing only one independent wooden board sample by a simple image processing method. By continuously using the network, multiple independent two-dimensional image samples can be generated without interruption, and that each image sample contains only one complete woodboard image data. In addition to this, the neural network can also perform the identification of the side boundaries and reject the irrelevant background image data.

In an alternative implementation of an embodiment of the present disclosure, image data may also be acquired using a plurality of linear cameras. Multiple linear cameras simultaneously acquire image data of a wooden board sample at different shooting angles and generate multiple sets of training data. The plurality of sets of training data include labels such as a board type, a corresponding speed, and a corresponding shooting angle, and may of course include a label such as a corresponding illumination. Multiple linear cameras can also use a variety of different lighting conditions. For example, the scanning area is divided into two areas, A and B, which are opaque, and linear cameras a and b are installed respectively, and s1 and s2 are used. After the board sample passes through the scanning area, two Samples under different lighting conditions are obtained. In the same way, samples under multiple parameters such as shooting angle and moving speed can also be obtained.

FIG. 7 illustrates a flow chart of a wood board identification method in accordance with an embodiment of the present disclosure. As shown in FIG. 7, the wood board identification method includes the following steps S701-S703:

In step S701, acquiring a plurality of one-dimensional images of the wood board;

In step S702, the acquired plurality of one-dimensional images are spliced to obtain a two-dimensional image to be identified;

In step S703, recognition is performed based on the two-dimensional image and the trained wood board recognition model, and the type of the board and the moving speed are obtained.

In this embodiment, the one-dimensional image can be acquired by a linear camera. Since the linear camera can only acquire one one-dimensional image at each time point, the one-dimensional image cannot be directly used for subsequent recognition. Therefore, a plurality of one-dimensional data can be spliced into one two-dimensional image, and the two-dimensional image can contain image information of a part or the whole board.

The wood board recognition model may be pre-trained, such as a wood board recognition model obtained using the machine learning method shown in FIG. Since the wood board recognition model is trained through two-dimensional images, speeds, and categories, the type of the board, the moving speed, and the like can be identified by the image samples of the board to be identified. The moving speed of the board is the relative moving speed, that is, the relative moving speed between the linear camera and the board that captures the image.

The wood board identification method can be performed in the control device of the wood board sorting system shown in FIG. As shown in FIG. 3, the board sorting system includes: a conveying device 301, a linear image collecting device 302, a control device 303, and a sorting device 304;

The wood to be classified is placed on the conveying device 301, and is conveyed backward by the conveying device 301;

The linear image capturing device 302 is disposed in alignment with the transmitting device 301 for collecting a one-dimensional linear image of the wood to be classified, and the output end of the linear image capturing device 302 is coupled to the control device;

The output end of the control device 303 is coupled to the classification device 304. The control device 303 outputs a direction signal and a time signal to the classification device 304 according to the one-dimensional image collected by the linear image acquisition device.

The sorting means 304 is disposed above the end of the transport means 301, and moves the board to be sorted out of the transport means 301 in the direction indicated by the direction signal at the point in time indicated by the time signal.

Alternatively, in one embodiment of the present disclosure, the transfer device is preferably a conveyor belt including, but not limited to, a belt, gear or chain driven conveyor belt device. As shown in FIG. 3, the transmitting device 301 can be further divided into an image capturing area 305 and a kicking area 306, wherein the linear image capturing device 302 is disposed in alignment with the image capturing area 305, and after the wood to be classified enters the image capturing area 305, The one-dimensional image of the wood to be classified is collected; the sorting device 304 is disposed in the kicking area 306 for moving the wood to be classified out to the designated position according to the direction signal and the time signal output by the control device 303.

In the embodiment of the present disclosure, the one-dimensional image collected by the linear image capturing device simultaneously performs the classification identification of the wood to be classified and the recognition of the kick machine control. On the one hand, at least one photoelectric sensor can be omitted, and on the other hand, the identification and control are performed. Speed and accuracy are higher than the relevant technology, greatly improving the efficiency of board sorting and reducing costs. Specifically, the linear image acquisition device collects a one-dimensional image of the wood to be classified in a relatively moving state under a certain natural and/or artificial illumination environment. Typically, as the board is transported on the conveyor, the board moves relative to the linear acquisition device with a relative speed between the two. The relative movement speeds in the present disclosure are not fixed, but may be variable, and are preferably varied. In the case where the relative moving speed is changed, the wood board sample images at different relative moving speeds can be better obtained.

Optionally, the linear image capture device is fixedly mounted above the transport device, and the linear image capture device is positioned to capture images of the image capture region of the transport device. During image acquisition, image acquisition at different speeds can be achieved by transforming the sampling frame rate of the linear image capture device; optionally, the wood board sorting system further includes at least one LED light source 307, and at least one LED light source 307 is disposed at the transmitting device 301. Above, the image acquisition area 306 is illuminated; preferably, at least one LED light source 307 is disposed adjacent to or integrated with the linear image acquisition device 302. Different illumination conditions can be realized by adjusting the brightness of the LED light source 307 or the illumination direction of the light source. The LED light source can be a flat type LED light to provide relatively uniform illumination. Further, the brightness of the light source can be set to be sequentially increased or decreased to obtain a product. Samples in different light conditions; different lighting conditions can also be achieved by transforming one or more of aperture size, source brightness, and illumination direction. For example, in an alternative embodiment, the illumination of the LED light is used to increase the basic brightness of the image acquisition to a satisfactory level, and at the same time, by varying the aperture size, the illumination level of the base brightness is obtained. Thus, by combining the two transformation methods, a plurality of illumination conditions within a satisfactory range can be obtained. In addition, during the image acquisition process, the angle of the image acquisition device can also be dynamically adjusted to capture the image of the wood sample at different angles. The combination of various transformations can be further performed in terms of time and/or order, that is, different transformation combinations can be performed in different sequences at different times to acquire images.

In an embodiment of the present disclosure, a reference image may also be set during image acquisition to assist in improving the accuracy of image recognition. For example, preferably, a reference object area is provided in the image acquisition area 306 of the transport device 301, and a reference object 308 is disposed in the reference object area; wherein the reference object area and the reference object 308 remain stationary (ie, do not move with the transport device); linear image The collecting device needs to ensure that the image of the wood board sample and the image of the reference object are collected at the same time. Preferably, the reference object has a white surface and the white reference object can be used to provide a standard reference for white balance, brightness or other image parameters.

In a preferred embodiment of the present disclosure, the linear image capture device may be one or more linear cameras (such as a plurality of linear cameras, etc.), simultaneously acquiring an image of a wooden board sample by one or more linear cameras, and generating a sample data. The sample data includes corresponding labels of illumination, speed, acquisition angle, etc., and the data collected by the plurality of sensors becomes a combination of multiple angle image data with respect to one image sensor.

Further, control device 303 is also coupled to one or more computer devices. In an embodiment of the present disclosure, the detection of the wood board classification may be done locally or in the cloud. Specifically, the collected image data is sent locally to the cloud, and the information that the cloud can provide includes, but is not limited to, the definition of the wood board classification, the sample image of each category, the classification recognition model, and the classification detection result.

In an optional implementation manner of the embodiment of the present disclosure, as shown in FIG. 8 , step S702 is a step of splicing the acquired one-dimensional images to obtain a two-dimensional image to be recognized, and further includes:

In step S801, the acquired plurality of one-dimensional images are divided into at least two sets of one-dimensional images according to an image acquisition time sequence and/or an image order;

In step S802, at least two sets of one-dimensional images are respectively spliced to form at least two two-dimensional images to be identified.

In the optional implementation manner, the wooden board is sent to the image collection area through the conveyor belt, and the wooden board completes the image acquisition through the scanning area of the linear camera during the movement, and processes the collected one-dimensional images through the linear camera to obtain the image. A two-dimensional image is entered and the acquired two-dimensional image is entered into a trained wood board recognition model. Alternatively, the board is fixed in an area, and the entire board is scanned by moving a linear camera to acquire a plurality of one-dimensional images and obtain a two-dimensional image.

In some cases, an external light source, such as an LED light source, can also be used during image acquisition. The light source provides a uniform illumination to enhance the underlying brightness of the image. At the same time, by using a linear camera synchronized with an external light source, different lighting conditions can be transformed at different times, and image samples under multiple illumination conditions can be obtained.

Then, in order to obtain a more accurate recognition result, the acquired one-dimensional one-dimensional image can be divided into two parts, that is, divided into two parts according to chronological order, and finally one-dimensional images of the front and back two parts are respectively spliced to obtain two one-dimensional images. . As before, when the first half of the one-dimensional image and the second half of the one-dimensional image are sampled, different illumination conditions can be used, so that the two two-dimensional images obtained have the same predetermined speed, but the illumination conditions are different.

Similarly, a set of one-dimensional images acquired at the same predetermined speed can be divided into two parts, that is, the one-dimensional images in which the sampling order of the obtained one-dimensional image is divided into odd and even digits are respectively divided into two groups, each group The included one-dimensional images are stitched into a two-dimensional image, and two two-dimensional images can be obtained at the same predetermined speed. As before, when sampling and acquiring a one-dimensional image at an odd time and sampling a one-dimensional image at an even time, different illumination conditions can be used, so that the two two-dimensional images obtained have the same predetermined speed, but the illumination conditions are different.

In an optional implementation manner of the embodiment of the present disclosure, in step S703, the two-dimensional image and the trained wood board recognition model are identified, and the category of the board and the moving speed are obtained, including:

In the optional implementation manner, when two two-dimensional images obtained by the wood board to be recognized are identified, two two-dimensional images to be recognized may be input to the wood board recognition model respectively, and the final recognition result is trusted. A higher-valued group is used as the final recognition result, which can increase the recognition accuracy.

The trained wood board recognition model analyzes the input two-dimensional image to be identified to determine the moving speed and category of the board. If, during training, each type of training sample contains training samples under different lighting conditions, the trained wood board recognition model can implement reliable speed and category recognition under any lighting conditions. Here, any lighting condition may be an illumination condition that floats up and down around the base brightness under a basic brightness condition, such as by illumination of an external source of the LED. If the previous sample collection does not contain multiple lighting conditions, but only the category identification, the recognition error may be caused by changes in lighting conditions. This is due to the lack of sample data under multi-light conditions, and the convolutional neural network cannot correct the effects of illumination on image samples. The color feature of the image is inevitably used in the recognition process, and the feature changes with the change of illumination, so that different illumination changes the final classification result. When image samples under multiple illumination conditions of the same wood sample are used, multiple sets of confidence estimates will be obtained, preferably with the result with the highest confidence value as the final result.

In addition, a reference image, such as a white reference image, can also be used during image acquisition. The image is captured into an image data simultaneously with the wood board image, which can be used as a reference for white balance and brightness or other image parameters.

At this time, the white reference described above can correct the image for white balance and brightness. Since the white reference can be considered as an image known, the change in illumination of the woodboard image can be derived from the change in illumination of the white reference. Since the speed of the board on the conveyor belt is not always equal to the speed of the conveyor belt, each new board can enter the image acquisition area at an arbitrary speed. Since the training samples contain samples at different speeds, the board identification model is also Can distinguish the speed of movement of the board. In an embodiment, the predetermined kick time may be determined according to the moving speed of the wooden board, and the kicking operation, that is, the sorting operation, is performed according to the determined wooden board type at the predetermined kicking time, and the wooden board is kicked into the corresponding category. .

An example neural network output is as follows:

Sample 1:

[Category: {A: 95%, B: 3%, C: 2%}, speed: {V0: 99%, V1:1%}, camera angle: {A1: 100%}]

Sample 2:

[Category: {A: 1%, B: 99%, C: 0%}, speed: {V0: 2%, V1: 98%}, camera angle: {A1: 100%}]

When a linear camera obtains multiple 2D images of the same board, the output of the board recognition model may be:

Sample 1a:

Sample 1b:

[Category: {A: 92%, B: 3%, C: 5%}, speed: {V0: 99%, V1: 98%}, camera angle: {A1: 100%}]

Sample 1a can be used for the final classification judgment at this time. This is due to the fact that Sample 1a is more suitable for illumination conditions due to natural light and external light source synthesis, thus making the confidence value higher.

That is, the output of the board recognition model is the final estimate of the confidence of each category (eg board type, speed, shooting angle, etc.). Based on these estimates, the category with the highest confidence can be selected as the final output. Note that the above implementation only uses neural networks as a basic method. Other similar machine learning methods, such as support vector machine, KNN, RNN, K-means, decision forest, etc., can continue the same method and process. Implement a solution based on other machine learning methods. .

The action performed by the kick can be based on a time-to-speed mapping relationship, for example:

T=aV+b

Where V is the resulting velocity estimate and a, b are predefined parameters. If the conveyor belt changes, such as when the distance between the kicker and the camera changes, only the values of a and b need to be changed without retraining the entire neural network. The mapping relationship between kick time and speed can also be achieved in many ways, not limited to methods.

In an optional implementation manner of the embodiment of the present disclosure, in step S703, the two-dimensional image and the trained wood board recognition model are identified, and the steps of the board type and the moving speed are obtained, including:

In this alternative implementation, the linear camera may acquire a one-dimensional image uninterruptedly on a plurality of planks that are moved or arranged in a pipeline, and spliced to obtain a continuous two-dimensional image. In this case, a trained cutting model, that is, a board boundary recognition model, can be used to cut a continuous two-dimensional image, that is, to identify boundary information in the two-dimensional image, and to divide the boundary information into multiple A two-dimensional image is obtained for each two-dimensional image sample containing only one wood board image. After receiving the uninterrupted two-dimensional image, the cutting model can generate a decision of the starting boundary at time t1 and generate a decision to end the boundary at time t2. Further, the image data at times t1 and t2 is composed into an image containing only one independent wooden board sample by a simple image processing method. By continuously using the network, multiple independent two-dimensional image samples can be generated without interruption, and that each image sample contains only one complete woodboard image data. In addition to this, the neural network can also perform the identification of the side boundaries and reject the irrelevant background image data.

The cut image data containing the individual boards is used as input, and an output value is generated when passing through each layer of the neural network, and is used for the input of the next layer. After passing through all levels, the convolutional neural network will get an estimate of confidence on each classification. If the linear camera obtains image samples under multiple illumination conditions, it can be input to the convolutional neural network separately to obtain multiple confidence evaluation groups.

In the actual identification process, since the board is a non-standardized product, the board identification model trained under certain conditions cannot be accurately classified. For example, the final confidence estimates of category A and category B are approximately equal, for example 51. % vs. 49%, the confidence of this classification is considered to be low. For such boards, the operator can be notified to intervene by dividing them into a single "unrecognizable classification" or by means of an alarm device. At this point, such boards can be processed in an iterative manner. First, it is classified by manual method, and then the image samples of the sample under multi-speed, multi-light, and multi-camera angles are obtained. This results in a new image sample that can be added to the sample used to train the neural network and retrain the neural network when the time is right. By continuously increasing the sample data and training the neural network, the processing accuracy of various abnormal samples can be continuously increased, and the probability of occurrence of low confidence is gradually reduced. An example of the data in the iterative process is given below:

Sample 1

Sample 2

...

.

Sample N

It is the sample data used to train the neural network for the first time. Through the iterative method, the following samples are obtained:

Sample 1

Sample 2

...

Sample N

Sample N+1

Sample N+2

Sample N+3

Among them, Sample N+1, Sample N+2, and Sample N+3 are all corresponding to the same low-confidence board, but correspond to different speed, illumination, camera angle and other attributes. Since the image samples may be affected by the material of the original wood, the image samples of different batches of logs are different. At the same time, even the same batch of logs may change the custom classification and the correspondence between the classification and the sample due to changes in coloring or demand. These rapidly changing needs can quickly adjust the sorting algorithm by changing image samples and retraining the neural network. E.g,

SampleSet 1, SampleSet2, SampleSet3.....

It is a different training data set, which corresponds to different raw wood quality, different classification methods, different painting processes and so on. As long as the corresponding training data is used, the neural network is quickly trained and the sorting algorithm is updated, the new requirements are met at any time without changing any hardware structure. Especially in the real-time situation of the training process in the cloud, the training process can be greatly shortened, so that the parameter adjustment, calibration, testing and deployment process that can be completed in a few months can be completed in one day.

The above-described wood board recognition method proposed by the present disclosure can accurately classify wood boards regardless of any environment and any speed, and perform classification operations at accurate kick time. Moreover, there is no need for a separate device, such as a photoelectric sensor, to obtain the time of arrival of the board, but the classification time can be determined by classification learning.

The following is an apparatus embodiment of the present disclosure, which may be used to implement the method embodiments of the present disclosure.

9 is a block diagram showing the structure of a machine learning device for wood board recognition according to an embodiment of the present disclosure, which may be implemented as part or all of an electronic device by software, hardware, or a combination of both. As shown in FIG. 9, the board learning device includes a first acquisition module 901, a first splicing module 902, and a training module 903:

The first obtaining module 901 is configured to acquire a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds; wherein each set of one-dimensional images includes a plurality of one-dimensional images at different positions of the corresponding wooden boards, and each set of one-dimensional Multiple one-dimensional images in the image correspond to the same predetermined speed;

The first splicing module 902 is configured to separately splicing each set of one-dimensional images in the acquired plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

The training module 903 is configured to train the board recognition model as the plurality of sets of training data by using the category of the board, the plurality of different predetermined speeds, and the plurality of two-dimensional images; each group of the training data includes the category of the board a two-dimensional image of the plurality of two-dimensional images and a corresponding predetermined speed; the recognition result of the wood board recognition model includes a category of the wood board and a moving speed.

In an optional implementation manner of this embodiment, the first acquiring module includes:

Training modules, including:

The first splicing module includes:

In an optional implementation manner of the embodiment, the illumination condition includes the intensity of the light of the external light source of the wooden board, the illumination direction of the light of the external light source, the shooting angle of the image acquiring unit that acquires the one-dimensional image, and the aperture size of the image acquiring unit. One or more; the predetermined speed is the relative moving speed between the image acquisition unit and the board.

In an optional implementation manner of this embodiment, the method further includes:

In an optional implementation manner of this embodiment, the first splicing module includes:

The third splicing sub-module is configured to splicing each set of one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images;

In an optional implementation manner of this embodiment, the training module further includes:

The machine learning device for the wood board recognition is consistent with the machine learning method for the wood board recognition. For details, refer to the above description of the machine learning method for wood board recognition, and details are not described herein.

FIG. 10 is a block diagram showing the structure of a wood board identifying apparatus according to an embodiment of the present disclosure, which may be implemented as part or all of an electronic device by software, hardware, or a combination of both. As shown in FIG. 10, the board identification device includes a third acquisition module 1001, a second splicing module 1002, and an identification module 1003:

The third obtaining module 1001 is configured to acquire a plurality of one-dimensional images of the wood board;

The second splicing module 1002 is configured to splicing the acquired one-dimensional images to obtain a two-dimensional image to be identified;

The identification module 1003, the splicing module identifies the two-dimensional image and the trained wood board recognition model, and obtains the category of the board and the moving speed.

In an optional implementation manner of this embodiment, the second splicing module includes:

In an optional implementation manner of this embodiment, the identifying module includes:

The device for identifying the wood board is consistent with the method for identifying the wood board. For details, refer to the description of the board identification, which is not described here.

As shown in FIG. 11, the electronic device 1100 includes a central processing unit (CPU) 1101 that can be loaded into a program in a random access memory (RAM) 1103 according to a program stored in a read only memory (ROM) 1102 or a program stored from the storage portion 1108. The various processes in the embodiment shown in Fig. 1 described above are executed. In the RAM 1103, various programs and data required for the operation of the electronic device 1100 are also stored. The CPU 1101, the ROM 1102, and the RAM 1103 are connected to each other through a bus 1104. An input/output (I/O) interface 1105 is also coupled to bus 1104.

The following components are connected to the I/O interface 1105: an input portion 1106 including a keyboard, a mouse, etc.; an output portion 1107 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker; a storage portion 1108 including a hard disk or the like And a communication portion 1109 including a network interface card such as a LAN card, a modem, or the like. The communication section 1109 performs communication processing via a network such as the Internet. Driver 1110 is also connected to I/O interface 1105 as needed. A removable medium 1111 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory or the like is mounted on the drive 1110 as needed so that a computer program read therefrom is installed into the storage portion 1108 as needed.

In particular, according to an embodiment of the present disclosure, the method described above with reference to FIG. 1 may be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product comprising a computer program tangibly embodied on a readable medium therewith, the computer program comprising program code for performing the machine learning method of the board recognition of FIG. In such an embodiment, the computer program can be downloaded and installed from the network via the communication portion 1109, and/or installed from the removable medium 1111.

The above electronic device can also be used to execute the program code of the wood board identification method in the embodiment shown in FIG.

Claims

A machine learning method for wood board recognition, which comprises:

Obtaining a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds; wherein each set of one-dimensional images includes a plurality of one-dimensional images corresponding to different positions of the wooden board, and the each of the sets of one-dimensional images Multiple one-dimensional images corresponding to the same predetermined speed;

Combining each of the acquired one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

The board identification type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as a plurality of sets of training data to respectively train the board recognition model; each of the plurality of sets of training data includes the a category of the board, a two-dimensional image of the plurality of two-dimensional images, and a corresponding predetermined speed; the recognition result of the board recognition model includes a category of the board and a moving speed.
The machine learning method for wood board recognition according to claim 1, wherein acquiring a plurality of sets of one-dimensional images of the wood board at a plurality of different predetermined speeds comprises:

Acquiring multiple sets of one-dimensional images of the wooden board in a case where the wooden board has a relative speed with a linear camera, and the combination of the relative speed and the sampling frame rate of the linear camera corresponds to the plurality of different predetermined speeds .
The machine learning method for wood board recognition according to claim 1, wherein acquiring a plurality of sets of one-dimensional images of the wood board at a plurality of different predetermined speeds comprises:

Obtaining a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds under different illumination conditions;

The board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as a plurality of sets of training data to respectively train the board recognition model, including:

The board identification type, the plurality of different predetermined speeds, the different illumination conditions, and the plurality of two-dimensional images are used as a plurality of sets of training data to respectively train the board recognition model; each of the plurality of sets of training data The group training data includes a category of the board, a two-dimensional image of the plurality of two-dimensional images, and corresponding pre-determined speeds and lighting conditions.
The machine learning method for wood board recognition according to claim 3, wherein each of the acquired one-dimensional images in the plurality of sets of one-dimensional images is separately spliced to obtain a plurality of two at a plurality of different predetermined speeds. Dimensional images, including:

Dividing each of the acquired one-dimensional images of the plurality of sets of one-dimensional images into at least two groups according to an image acquisition time sequence and/or an image order;

The one-dimensional images of the at least two groups are separately spliced to form a plurality of two-dimensional images at the different predetermined speeds.
The wood board recognition machine learning method according to claim 4, wherein the illumination condition comprises a light intensity of a light source external to the wood board, an illumination direction of the external light source light, a photographing angle of the image acquisition unit that acquires the one-dimensional image, and One or more of the aperture sizes of the image acquisition unit; the predetermined speed is a relative movement speed between the image acquisition unit and the board.
The machine learning method for wood board recognition according to claim 1, further comprising:

A plurality of one-dimensional images of white reference objects are also acquired while acquiring a plurality of sets of one-dimensional images of the wooden board.
The wood board recognition machine learning method according to claim 1, wherein the board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as a plurality of sets of training data to respectively perform a wood board recognition model Training, including:

Splicing each set of one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images;

The plurality of two-dimensional images are respectively labeled to obtain boundary information of the wooden board.
The wood board recognition machine learning method according to claim 7, wherein the board type, the plurality of different predetermined speeds, and the plurality of two-dimensional images are used as a plurality of sets of training data to respectively perform a wood board recognition model Training also includes:

The board boundary recognition model is trained according to the plurality of two-dimensional images and boundary information of the board, and the recognition result of the board boundary recognition model includes boundary information of the board.
A method for identifying a wood board, comprising:

Obtaining a plurality of one-dimensional images of the wooden board;

Splicing the acquired one-dimensional images to obtain a two-dimensional image to be identified;

Identification is performed based on the two-dimensional image and the trained wood board recognition model, and the type of the board and the moving speed are obtained.
The method for identifying a board according to claim 9, wherein the merging the acquired one-dimensional images to obtain a two-dimensional image to be recognized comprises:

And dividing the acquired plurality of one-dimensional images into at least two sets of one-dimensional images according to an image acquisition time sequence and/or an image order;

The at least two sets of one-dimensional images are separately spliced to form at least two two-dimensional images to be identified.
The board identifying method according to claim 10, wherein the identifying according to the two-dimensional image and the trained wood board recognition model, the category of the board and the moving speed are obtained, including:

Inputting the at least two two-dimensional images into the wood board recognition model respectively, to obtain two types of confidence estimates of the category of the board and the moving speed;

The group with the highest confidence is selected from the two sets of confidence estimates as the final recognition result.
The method for identifying a wood board according to claim 9, further comprising:

The kick timing of the board is obtained according to the moving speed of the board.
The board identifying method according to claim 9, wherein the two-dimensional image and the trained wood board recognition model are identified to obtain a category of the board and a moving speed, including:

Obtaining boundary information of the wooden board according to the two-dimensional image and the trained boundary recognition model;

According to the two-dimensional image, the boundary information, and the wood board recognition model, the category of the board and the moving speed are obtained.
The board identifying method according to claim 9, wherein the identifying according to the two-dimensional image and the trained wood board recognition model, the category of the board and the moving speed are obtained, including:

Inputting a plurality of the two-dimensional images obtained under a plurality of different illumination conditions into the wood board recognition model, respectively, to obtain a plurality of sets of confidence estimates of the category of the board and the moving speed;

A group with the highest degree of confidence is selected from the plurality of sets of confidence estimates as a final recognition result.
The method for identifying and identifying a wood board according to claim 9, further comprising:

A plurality of one-dimensional images of white reference objects are also acquired while acquiring a plurality of one-dimensional images of the wooden board.
A machine learning device for wood board recognition, comprising:

a first obtaining module configured to acquire a plurality of sets of one-dimensional images of the wooden board at a plurality of different predetermined speeds; wherein each set of one-dimensional images includes a plurality of one-dimensional images corresponding to different positions of the wooden board, and each The plurality of one-dimensional images in the set one-dimensional image correspond to the same predetermined speed;

a first splicing module configured to splicing each of the acquired one-dimensional images of the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images at a plurality of different predetermined speeds;

a training module configured to train the board recognition model as the plurality of sets of training data, the category of the board, the plurality of different predetermined speeds, and the plurality of two-dimensional images; each of the plurality of sets of training data The group training data includes a category of the board, a two-dimensional image of the plurality of two-dimensional images, and a corresponding predetermined speed; the recognition result of the board recognition model includes a category of the board and a moving speed.
The board-recognized machine learning apparatus according to claim 16, wherein the first obtaining module comprises:

a first acquisition submodule configured to acquire when the wood board has a relative speed with the linear camera, and the combination of the relative speed and the sampling frame rate of the linear camera corresponds to the plurality of different predetermined speeds Multiple sets of one-dimensional images of the wood board.
The board-recognized machine learning apparatus according to claim 16, wherein the first obtaining module comprises:

a second obtaining submodule configured to acquire a plurality of sets of one-dimensional images of the board at a plurality of different predetermined speeds under different lighting conditions;

The training module includes:

a first training sub-module configured to train the board recognition model by using the category of the board, the plurality of different predetermined speeds, the different illumination conditions, and the plurality of two-dimensional images as sets of training data Each of the plurality of sets of training data includes a category of the board, a two-dimensional image of the plurality of two-dimensional images, and corresponding predetermined speeds and lighting conditions.
The machine tool for wood board recognition according to claim 16, wherein the first splicing module comprises:

a first grouping sub-module configured to divide each of the acquired one-dimensional images of the plurality of sets of one-dimensional images into at least two groups according to an image acquisition time sequence and/or an image order;

The first splicing sub-module is configured to splicing the one-dimensional images of the at least two groups separately to form a plurality of two-dimensional images at the different predetermined speeds.
The board-recognized machine learning apparatus according to claim 18, wherein said illumination condition comprises a light intensity of a light source external to the wooden board, an irradiation direction of the external light source light, a photographing angle of the image acquisition unit that acquires the one-dimensional image, and One or more of the aperture sizes of the image acquisition unit; the predetermined speed is a relative movement speed between the image acquisition unit and the board.
A wood board-recognized machine learning apparatus according to claim 16, further comprising:

The second obtaining module is configured to acquire a plurality of one-dimensional images of the white reference object while acquiring the plurality of sets of one-dimensional images of the wooden board.
The machine tool for wood board recognition according to claim 16, wherein the first splicing module comprises:

a second splicing sub-module configured to splicing each set of one-dimensional images in the plurality of sets of one-dimensional images to obtain a plurality of two-dimensional images;

The labeling sub-module is configured to label the plurality of two-dimensional images separately to obtain boundary information of the wood board.
The board-recognized machine learning apparatus according to claim 22, wherein the training module further comprises:

The second training sub-module is configured to train the board boundary recognition model according to the plurality of two-dimensional images and boundary information of the board, the recognition result of the board boundary recognition model including boundary information of the board.
A wood board identification device, comprising:

a third obtaining module configured to acquire a plurality of one-dimensional images of the wood board;

a second splicing module configured to splicing the acquired one-dimensional images to obtain a two-dimensional image to be identified;

The identification module, the splicing module performs recognition according to the two-dimensional image and the trained wood board recognition model, and obtains the category of the wood board and the moving speed.
The wood board identification device according to claim 24, wherein the second splicing module comprises:

a second grouping sub-module configured to divide the acquired plurality of one-dimensional images into at least two sets of one-dimensional images according to an image acquisition time sequence and/or an image order;

The third splicing sub-module is configured to splicing the at least two sets of one-dimensional images separately to form at least two two-dimensional images to be identified.
The wood board identification device according to claim 25, wherein the identification module comprises:

a first identification sub-module configured to input the at least two two-dimensional images into the wood board recognition model respectively, to obtain two types of confidence estimates of the category of the board and the moving speed;

The first selection sub-module is configured to select the group with the highest confidence from the two sets of confidence estimates as the final recognition result.
The wood board identification device according to claim 24, further comprising:

The kick leg module is configured to obtain a kick timing of the board according to a moving speed of the board.
The wood board identification device according to claim 24, wherein the identification module comprises:

a second identification sub-module configured to obtain boundary information of the wooden board according to the two-dimensional image and the trained boundary recognition model;

The third identification sub-module is configured to obtain a category of the wooden board and a moving speed according to the two-dimensional image, the boundary information, and the wood board recognition model.
The wood board identification device according to claim 24, wherein the identification module comprises:

a fourth identification sub-module configured to input a plurality of the two-dimensional images obtained under a plurality of different illumination conditions to the wood board recognition model respectively, to obtain a plurality of confidence estimates of the type of the board and the moving speed ;

The second selection sub-module is configured to select the group with the highest confidence from the plurality of sets of confidence estimates as the final recognition result.
A wood board-recognized machine learning apparatus according to claim 24, further comprising:

The fourth obtaining module is configured to acquire a plurality of one-dimensional images of the white reference object while acquiring the plurality of one-dimensional images of the wood board.
An electronic device, comprising a memory and a processor; wherein

The memory is for storing one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method steps of any of claims 1-8.
An electronic device, comprising a memory and a processor; wherein

The memory is for storing one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method steps of any of claims 9-15.
A computer readable storage medium having stored thereon computer instructions, wherein the computer instructions are executed by a processor to implement the method steps of any of claims 1-8.
A computer readable storage medium having stored thereon computer instructions, wherein the computer instructions, when executed by a processor, implement the method steps of any of claims 9-15.