TWI786555B - Pattern identification and classification method and system - Google Patents

Abstract

A pattern recognition and classification method is applicable to a computing device, and includes the following steps: (A) taking plural recognition images from an analysis image related to the pattern. (B) A neural network is used, and the neural network outputs a set of feature values according to each recognized image, and each set of feature values has a plurality of feature values. (C) comparing each of the characteristic value groups with a standard characteristic value group, and judging the classification of the plate according to the comparison result, the standard characteristic value group having a plurality of standard characteristic values respectively corresponding to the characteristic values. In this way, not only can the accuracy of identification and classification be greatly improved, but also the time and cost of updating and maintaining identification software can be greatly reduced.

Description

Pattern identification and classification method and system

The present invention relates to an identification and classification method and system, in particular to a plate identification and classification method and system.

In the current automated production line of furniture materials, the panels used as furniture materials need to be identified and classified before they leave the factory to ensure that the products shipped are correct.

Since the pattern usually has a large area, for example, it is usually a pattern with a size of more than 2 meters x 2 meters. Therefore, the photographic device will be used to divide the pattern into multiple images, and then the identification program will be used to identify the Image recognition and classification. The current identification program construction method is to manually list the identification features of wood grains, such as the distribution of tones, specific curves, specific shapes, etc. of walnut wood grains, and then manually write identification calculations based on these identification features. formula, and then identify and classify the images by using a recognition program including the recognition calculation formula.

However, this identification method is limited by the limited number of wood grain identification algorithms that can be listed manually. Not only does the identification accuracy still have room for improvement, but it is also difficult to quickly expand and update the identification program as the variety of boards increases.

Therefore, the object of the present invention is to provide a method for identifying and classifying plates that can solve the above-mentioned problems.

Therefore, the pattern identification and classification method of the present invention is applicable to a computing device, and includes the following steps:

(A) Taking a plurality of identification images in an analysis image related to the plate.

(B) A neural network is used, and the neural network outputs a set of feature values according to each recognized image, and each set of feature values has a plurality of feature values.

(C) comparing each of the characteristic value groups with a standard characteristic value group, and judging the classification of the plate according to the comparison result, the standard characteristic value group having a plurality of standard characteristic values respectively corresponding to the characteristic values.

Therefore, the object of the present invention is to provide a plate identification and classification system that can solve the above-mentioned problems.

Therefore, the plate identification and classification system of the present invention includes a conveying device, a photographing device, a lighting device, and a computing device.

The conveying device is used for conveying the plate.

The photographing device is set corresponding to the conveying device for photographing the plate and outputting a plurality of partitioned images.

The illuminating device is arranged corresponding to the photographing device, and is used for providing illumination when the photographing device photographs the plate.

The computing device is signal-connected to the photographing device, and the analysis image is obtained according to the partitioned images to execute the steps of the above-mentioned method.

The effect of the present invention is: by taking the identification images from the analysis image related to the plate, and then using the neural network to calculate the identification images to obtain the feature value groups, and then using the features The value group is compared with the standard feature value group to determine the classification of the pattern. Not only can the accuracy of identification and classification be greatly improved, but also the time and cost of updating and maintaining identification software can be greatly reduced.

Referring to FIG. 1 and FIG. 2 , an embodiment of the plate identification and classification system of the present invention includes a conveying device 2 , a photographing device 3 , a lighting device 4 , and a computing device 5 .

The conveying device 2 is used to convey the sheet 9 forward. The conveying device 2 can be implemented using crawler belts or a plurality of rollers that rotate in parallel.

The photographing device 3 is set corresponding to the conveying device 2 for photographing the plate 9 and outputting a plurality of partitioned images 71 . The imaging device 3 is a device capable of capturing images, for example, using a CMOS color camera, a CCD color camera, etc., wherein it is preferable to use a CCD camera with better sensitivity, resolution, and noise control to obtain a better image quality. image quality. Wherein, because the area of the plate 9 of general furniture is relatively large, it is usually impossible to accommodate the entire plate 9 in one image, so the plate 9 will be photographed in partitions to obtain a plurality of pictures that can be combined It is a subregional image 71 of an overall image, and when the area of the plate 9 is small, a subregional image 71 can be used as the overall image for subsequent analysis. Wherein, the size of each partition image 71 is preferably between 20 cm x 20 cm ~ 60 cm x 45 cm.

The lighting device 4 is arranged corresponding to the photographing device 3 to provide illumination for the photographing device 3 to photograph the plate 9 . In order to reduce color shift and improve the accuracy of wood grain color, the lighting device 4 is preferably a white light fixture, for example, a light fixture with a color temperature between 4500K~6500K, preferably a pure white light fixture with a color temperature of about 6000K.

The computing device 5 is signal-connected to the photographing device 3, receives the subregional images 71, and obtains an analysis image 7 based on the subregional images 71, and uses them to execute the program instructions in the internal memory to implement a method for identifying and classifying plates, and by The plate 9 is identified and classified by the plate identification and classification method. Wherein, the computing device 5 and the photographing device 3 can use a wired connection or a wireless connection (such as Wi-Fi, Bluetooth and other wireless transmission technologies) to transmit the partitioned images 71 . The computing device 5 can use electronic devices with image processing computing capabilities such as computers and servers, for example, can use graphics processing units (Graphics Processing Unit, abbreviated as GPU), neural network processors (Neural Network Processing Unit, abbreviated NPU), or tensor processing unit (Tensor Processing Unit, abbreviated as TPU) computer implementation.

Referring to Fig. 1, Fig. 2, Fig. 3 and Fig. 4, one embodiment of the pattern identification and classification method of the present invention is applied to the computing device 5, and is applicable to the above-mentioned pattern identification and classification system. The pattern identification and classification method includes the following step:

Step 81 : The computing device 5 obtains a complex identification image 711 from the analysis image 7 related to the plate 9 .

Wherein, the analysis image 7 is an overall image formed by combining a plurality of partition images 71 shot toward the plate 9 , and the identification images 711 are preferably evenly distributed and taken from the analysis image 7 in an odd number. For example, as shown in FIG. 2 , one identification image 711 is taken from each of the four corners of the analysis image 7 , and one identification image 711 is taken from the center of the partition images 71 in the middle to obtain seven identification images 711 . In the case of both image recognition accuracy and computing load, the number of these recognition images 711 is preferably less than 20, for example, 5, 7 or 9 can be used, and the size of each recognition image 711 is preferably Between 8 cm x 8 cm ~ 12 cm x 12 cm, it is better to use a size of about 10 cm x 10 cm.

Step 82: The computing device 5 uses a neural network 6 (Neural Network, NN) in memory, and the neural network 6 outputs a feature value group 61 according to each identification image 711, and each feature value group 61 has complex feature values 611.

Wherein, each eigenvalue group 61 preferably has more than 500 eigenvalues 611 corresponding to the characteristics of the plate 9 . For example, there may be 500, 600, 1000 or more than 10000 feature values 611, each feature value 611 is a feature corresponding to a kind of wood grain, for example, these feature values 611 may be the red tone ( R) distribution state, green tone (G) distribution state, blue tone (B) distribution state, line density, line curvature distribution state, etc. Since these eigenvalues 611 are calculated by the eigenvectors extracted by the neural network 6 during the training process, there is no need to manually list one by one and establish a corresponding identification program.

The neural network 6 can be implemented using Convolutional Neural Networks (CNN for short), which are currently widely used in the field of image recognition, and the convolutional layer (convolutional layer), linear The rectification layer (rectified linear units layer, abbreviated as ReLU layer), the number of pooling layers is established, and the size of the convolution kernel (Kernel) in each convolution layer is set.

Step 83: compare these feature value groups 61 with a standard feature value group 62 respectively, and judge the classification of the plate 9 according to the comparison result, the standard feature value group 62 has a plurality of standards corresponding to the feature values 611 respectively Eigenvalue 621. The number of the standard feature values 621 in the standard feature value group 62 is the same as the number of the feature values 611 in each feature value group 61 .

Wherein, Euclidean distance (or called Euclidean distance) or cosine distance (or cosine similarity) can be used to calculate the difference value between each feature value group 61 and the standard feature value group 62 as a comparison, and in most features When the difference values between the value group 61 and the standard characteristic value group 62 are both smaller than a standard predetermined value, it is judged that the plate 9 is the wood grain corresponding to the standard characteristic value group 62 . For example, assuming that the neural network 6 outputs 7 sets of feature value groups 61 according to 7 identification images 711, and each feature value group 61 has 1000 feature values 611, the calculation of the 1000 feature values 611 of each feature value group 61 and The Euclidean distance of the 1000 standard eigenvalues 621 of the standard eigenvalue group 62, when the calculated Euclidean distance is less than the standard predetermined value, it is judged that the identification image 711 corresponding to the group of eigenvalues 61 belongs to the If the wood grain corresponding to the standard feature value group 62 is greater than the standard predetermined value, then it is judged that the identification image 711 corresponding to the feature value group 61 does not belong to the wood grain corresponding to the standard feature value group 62, and use Majority voting (voting) method determines the final judgment result. That is, when a kind of woodgrain is used as the comparison reference, when more than half of the 7 recognized images 711 are judged to be smaller than the predetermined standard value, the final judgment result is that the kind of woodgrain is met. If the comparison standard is multiple wood grains, assuming that among the 7 recognition images 711, 3 recognition images 711 are judged to be walnut, 2 recognition images 711 are judged to be birch, and 2 recognition images 711 are judged to be cherry wood, then finally Judgment result is most walnut.

Referring to Fig. 1, Fig. 2, Fig. 4 and Fig. 5, the method for identifying and classifying plates preferably also includes the following steps for training (training) the neural network 6:

Step 84: The computing device 5 receives a plurality of subregional images 71 shot toward the plate 9, and obtains a plurality of training images from each subregional image 71.

Wherein, in the training process, in order to improve the calculation accuracy of the neural network 6, it is preferable to input a larger number of these training images for training, for example, these training images can be extracted according to the size of the partition image 71 , when the subregional image 71 is larger, more training images are captured, the number of which is adjusted between 30 and 10000 according to the size of the subregional image 71 . The principle of selecting these training images is to set a plurality of target detection points in each subregion image 71, and repeatedly capture 20-50 different positions at each target detection point (that is, the positions will have displacements up, down, left, and right. ) of the training images, the repetition area of each training image is preferably no more than 70%. The target detection point can be each pixel, or a plurality of detection points with preset positions.

In addition, the subregional images 71 preferably include image information of the plate 9 taken from different angles, or different illuminance, or different distances, or select image information including a combination of the above-mentioned situations according to requirements, for training The neural network 6 can correctly judge wood grain under various conditions. Among them, the different angles referred to are to take pictures after rotating the plate 9 horizontally at 0 degrees, 90 degrees, 180 degrees, and 270 degrees; the different illuminances referred to are the average illuminance value during factory inspection, and the average illuminance of ±5%. value, the average illuminance value of ±10% to photograph the plate 9; the different distances referred to are the average distance when the photographic device 3 is used to photograph the plate 9 when the ex-factory inspection is used, the average distance of ±5%, the average distance of ±5%. The plate 9 is photographed at an average distance of 10%. In this way, image information of different reduction ratios can be obtained through different photographing distances.

The size of each training image is the same as that of the recognition image 711 , preferably between 8 cm x 8 cm ~ 12 cm x 12 cm, more preferably about 10 cm x 10 cm.

Step 85: The neural network 6 outputs the corresponding feature value set 61 according to each training image.

Same as above, each eigenvalue group 61 has more than 500 eigenvalues 611 corresponding to the characteristics of the plate 9, and each eigenvalue 611 is a characteristic corresponding to a kind of wood grain. These eigenvalues 611 are obtained through operation on the eigenvectors extracted by the neural network 6 during the training process.

Step 86: Adjust the parameters of the neural network 6 to reduce the loss between the feature value groups 61 output by the neural network 6 according to the training images.

Among them, the parameters of the neural network 6 can be adjusted by using the contrastive loss method (Contrastive loss), so that the distance between the feature value groups 61 between the same wood grains becomes smaller, and the distance between the feature value groups 61 between different wood grains become larger (training images with different wood grains need to be prepared for training together), or use the triplet loss method (triplet loss) or the improved triplet loss method (Improved triplet loss) to adjust the parameters of the neural network 6, Make the distance between the feature value groups 61 between the same wood grains smaller, and the distance between the feature value groups 61 between different wood grains become larger (training images with different wood grains need to be prepared and trained together). When the distance between the feature value groups 61 between the same wood grains is less than a predetermined adjustment value, it means that the neural network 6 has reached an acceptable recognition accuracy. The predetermined adjustment value can be adjusted according to the actual required accuracy.

Step 87: Obtain the standard feature value set 62 according to the feature value sets 61 corresponding to the training images.

Wherein, the average value of the feature value groups 61 corresponding to the training images is taken as the standard feature value group 62 . That is, the average value of the feature value groups 61 corresponding to the same wood grain is taken as the standard feature value group 62 . In this way, after using a large number of training images of various wood grains to train the neural network 6, a complex set of standard feature values 62 corresponding to various wood grains will be obtained for wood grain identification in steps 81-83. Classification.

Through the above description, the effect of this embodiment is as follows:

1. By taking the identification images 711 from the analysis image 7 related to the plate 9, and then using the neural network 6 to calculate the identification images 711 to obtain the feature value groups 61, and then by the The equal characteristic value group 61 is compared with the standard characteristic value group 62, and the classification of the plate 9 is judged accordingly. Not only can the neural network 6 be able to extract the characteristics of these eigenvalues 611 during the training process, but also construct a large number of eigenvalues 611 that cannot be achieved by manpower, thereby greatly improving the accuracy of identification and classification, and, when the product When the types of boards to be identified online increase, it is only necessary to retrain the neural network 6 to use it, without having to use manual reconstruction and expansion of new identification programs. Therefore, the time for updating and maintaining identification software can also be greatly reduced and cost.

To further illustrate, technically speaking, after training using, for example, the contrastive loss method (Contrastive loss), the triplet loss method (triplet loss), and the improved triplet loss method (Improved triplet loss), the resulting neural network The eigenvalue groups 61 output by road 6 for different wood grains will have certain differences. Therefore, it is not necessary to retrain the neural network 6 every time a board type is added, and the original neural network 6 can be used for After calculating the array eigenvalue group 61 for the new plate type and taking the average value to obtain the new standard eigenvalue group 62, the original neural network can be used without retraining the neural network 6 Lu 6 identifies new board types. However, the neural network 6 can also be retrained every time one or several board types are added, so that higher identification accuracy can be achieved. Wherein, if the wood grain of the newly added board has a high similarity with the original wood grain, it is recommended to retrain the neural network 6 to achieve better recognition accuracy.

2. By combining the subregional images 71 taken toward the plate 9 into the analysis image 7, and then extracting the identification images 711 from the analysis image 7 of the overall image, it is possible to reduce the problem of only considering a single part The error caused by the partition image 71.

3. By using the image information of the plate 9 taken from different angles, or different illuminances, or different distances for training, the recognition accuracy of the neural network 6 under various changes can be improved, and it is avoided when the production line When the environment changes slightly, the neural network 6 will produce misjudgment.

To sum up, the plate recognition and classification method and system of the present invention can indeed achieve the purpose of the present invention.

But what is described above is only an embodiment of the present invention, and should not limit the scope of the present invention. All simple equivalent changes and modifications made according to the patent scope of the present invention and the content of the patent specification are still within the scope of the present invention. Within the scope covered by the patent of the present invention.

2: Conveyor 3: Photographic device 4: Lighting device 5: Computing device 6: Neural Network 61:Eigenvalue group 611:Eigenvalue 62:Standard eigenvalue group 621:Standard eigenvalues 7: Analyzing images 71: Partition image 711: Recognition image 81~87: Steps 9: Plate

Other features and effects of the present invention will be clearly presented in the implementation manner with reference to the drawings, wherein: Fig. 1 is a schematic diagram of an embodiment of the plate identification and classification system of the present invention; Fig. 2 is a schematic diagram illustrating an analysis image, multiple partition images and multiple identification images of the embodiment; Fig. 3 is a flowchart of an embodiment of the plate identification and classification method of the present invention; Figure 4 is a schematic diagram of the operation of the embodiment; and FIG. 5 is a flowchart of training a neural network in this embodiment.

6: Neural Network 61:Eigenvalue group 611:Eigenvalue 62:Standard eigenvalue group 621:Standard eigenvalues 711: Recognition image

Claims (9)

A method for identifying and classifying a plate, which is applicable to a computing device, and includes the following steps: (A) taking a plurality of identification images from an analysis image related to the plate, and the analysis image is a plurality of partitions taken towards the plate Combining images, the size of each subregion image is between 20 cm x 20 cm ~ 60 cm x 45 cm, and the size of each identification image is between 8 cm x 8 cm ~ 12 cm x 12 cm; (B) use a neural network that outputs a feature value set based on each identified image, each feature value set having a plurality of feature values; and (C) comparing each of the feature value sets with at least one standard feature value set, And judging the classification of the plate according to the comparison result, the at least one standard feature value group has a plurality of standard feature values corresponding to the feature values respectively. The pattern recognition and classification method as described in Claim 1, wherein, in step (A), N recognition images are taken from the analysis image, and N is an odd number and less than 20. The pattern identification and classification method according to claim 1, wherein, in step (B), each feature value group has more than 500 feature values corresponding to the features of the pattern. The plate identification and classification method as described in claim 1, wherein, in step (C), the difference value between each eigenvalue group and the at least one standard eigenvalue group is calculated using Euclidean distance or cosine distance, and in most When the difference values between the eigenvalue group and the at least one standard eigenvalue group are less than a standard predetermined value, the judgment Determining the plate is the texture corresponding to the at least one standard feature value group. The plate identification and classification method as described in claim 1, further comprising the following steps for training the neural network: (D) receiving a plurality of partitioned images taken towards the plate, and taking a plurality of training images in each partitioned image ; (E) the neural network outputs the corresponding set of feature values according to each training image; (F) adjusts the parameters of the neural network so that the neural network outputs the set of feature values according to the training images and (G) deriving the at least one standard feature value set according to the feature value sets corresponding to the training images. The method for identifying and classifying plates as described in Claim 5, wherein, in step (D), the partition images include image information of the plates taken from different angles, or different illuminances, or different distances. The pattern recognition and classification method according to Claim 5, wherein in step (F), the parameters of the neural network are adjusted using a contrastive loss method or a triplet loss method. A pattern recognition and classification method is applicable to a computing device, and includes the following steps: (A) extracting plural recognition images from an analysis image related to the pattern; (B) using a neural network, the neural network Outputting a set of feature values according to each identified image, each set of feature values having a plurality of feature values; and (C) comparing each set of feature values with at least one set of standard feature values Compare, and judge the classification of the plate according to the comparison result, the at least one standard feature value group has a plurality of standard feature values corresponding to the feature values respectively; the plate identification and classification method also includes the following for training the neural network Steps of: (D) receiving a plurality of subregional images taken towards the plate, and taking multiple training images in each subregional image, wherein a plurality of target detection points are set in each subregional image, and each target detection Pick up 20-50 training images with different positions, and the repetition area of each training image is not more than 70%; (E) the neural network outputs the corresponding feature value group according to each training image; (F) adjusting the parameters of the neural network so that the loss between the feature value groups output by the neural network according to the training images is reduced; and (G) obtaining the at least one standard eigenvalue set. A plate identification and classification system, comprising: a conveying device, used to transport the plate; a photographing device, installed corresponding to the conveying device, to shoot towards the plate, and output a plurality of partition images; a lighting device, corresponding to the The photographing device is set up to provide illumination when the photographing device photographs the plate; and a computing device is connected to the photographing device to obtain an analysis image based on the images of the subregions, and is used to execute requirements 1 to 8 Any one steps of the method described.

Images