Feature fusion classification method for multiple types of packaging bag
By using a feature fusion classification method that combines support vector machines, template matching, and neural networks, the problem of high efficiency and high accuracy in packaging bag classification has been solved, resulting in reduced misidentification, improved automated production efficiency, and enhanced market competitiveness of the factory.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2026-04-02
AI Technical Summary
Existing packaging bag classification technologies cannot meet the requirements of efficient and high-precision detection. Traditional algorithms are fast but have poor robustness, deep learning methods are slow and do not meet the requirements of timely classification, and random forest methods process samples independently for each decision tree, resulting in insufficient system robustness.
We employ a feature fusion classification method based on the idea of random forest, combining support vector machine, template matching and neural network. Through image acquisition, processing, enhancement, support vector machine model construction, template matching and neural network training, we finally classify the data through a majority voting mechanism.
It improves the accuracy of packaging bag classification, reduces misidentification, enhances the factory's automation and intelligence level and production efficiency, and strengthens product quality and market competitiveness.
Smart Images
Figure CN2024120956_02042026_PF_FP_ABST
Abstract
Description
Feature fusion classification method for multi-type packaging bags TECHNICAL FIELD
[0001] The application belongs to the technical field of multi-type packaging bag classification, and particularly relates to a feature fusion classification method for multi-type packaging bags. BACKGROUND
[0002] In factory automation production, timely classification of multi-type packaging bags plays an important role. Through classification of packaging bags, a factory can improve automation intelligence level and production efficiency, thereby improving product quality and market competitiveness of the factory.
[0003] Existing packaging bag classification technology lags behind and cannot meet the requirements of efficient and high-precision detection, affecting the accuracy and efficiency of detection. The traditional algorithm corresponding to the existing packaging bag classification technology has high speed but poor robustness and is easily disturbed. The deep learning method has slow speed and cannot meet the requirement of timely classification. Random forest is a commonly used method of deep learning, which is a classifier composed of multiple decision trees, but there is no association between each decision tree. When a new input sample enters the forest, each decision tree classifies the sample, and then the final classification result is obtained through voting or averaging. Random forest has high accuracy, strong interpretability, and good performance in many fields including classification, regression, and feature selection, and is a powerful machine learning algorithm.
[0004] To reduce classification errors and improve system robustness, the technical solution proposes a feature fusion classification method for multi-type packaging bags based on the random forest idea, which obtains the final result through the voting mechanism of minority submission to majority.
[0005] SUMMARY
[0006] The application provides a feature fusion classification method for multi-type packaging bags, which solves the above problems.
[0007] To solve the above technical problems, the application is implemented by the following technical solutions:
[0008] The classification robot to which the feature fusion classification method is applied includes a packaging bag, a conveyor belt, a camera module, a mechanical arm, and a stacking table. The packaging bag is transported to the camera module through the conveyor belt, image acquisition of the packaging bag to be detected is performed, then the classification module is used to classify the image collected by the camera module to determine the final category, and finally the mechanical arm places the packaging bag on the stacking table corresponding to the category. The classification module stores a computer program of the feature fusion classification method for multi-type packaging bags.
[0009] The feature fusion classification method for multi-type packaging bags includes the following steps:
[0010] S1, the conveyor belt transports the packaging bag to the camera module, and the camera module installed on the packaging machine collects images of the packaging bag to be detected;
[0011] S2, the image denoising module is used to process the collected packaging bag image, improve the image quality, and reduce interference;
[0012] S3, the image enhancement module is used to enhance the packaging bag image;
[0013] S4, a support vector machine model is constructed and solved; this step includes the following sub-steps:
[0014] S41, collect a plurality of packaging bag images, take a part to construct a sample training set, and take another part to construct a test set;
[0015] S42, a support vector machine model is constructed, and the training set is used to train the support vector machine model;
[0016] Wherein, the support vector machine model includes the following steps:
[0017] Establish the objective function of the shortest distance between the feature points and the hyperplane s.t.y i (w T x i +b)≥1,i=1,2…n. (2)
[0018] Wherein, w represents an n-dimensional vector, b represents a real number, s.t. represents a constraint condition; y i represents the label of the i-th sample, represents 1 or -1, respectively representing positive samples and negative samples; x i represents the feature vector of the i-th sample, which is an n-dimensional vector;
[0019] To solve the objective function, the Lagrange function is introduced, and the Lagrange function of the above formula can be written as:
[0020] Wherein, α=(α1+α2,…,α n ) T , w represents the normal vector of the dividing hyperplane, x i represents the i-th feature attribute of a single sample, and b represents the distance between the origin and the hyperplane;
[0021] Analyze the corresponding dual problem:
[0022] Wherein, α i and α jis the Lagrange multiplier, used to solve the dual problem, is a vector, each element corresponds to a sample respectively; y i and y j represent the labels of the i-th and j-th samples, indicating 1 or -1, representing positive samples and negative samples respectively; x i T represents the feature vector of the i-th sample, is an n-dimensional vector, T represents transposition; x j : the feature vector of the j-th sample, is also an n-dimensional vector;
[0023] The optimal solution is obtained
[0024] The optimal classification function is obtained:
[0025] Where, Sgn() represents a sign function, outputting 1 when the input is greater than 0, outputting -1 when the input is less than 0, and outputting 0 when the input is equal to 0; f(x) represents a classification function, used to calculate which category the sample x belongs to;
[0026] The support vector machine model uses a linear kernel function, the formula is as follows: K(x i ,x j )=(x i *x j ) (7)
[0027] Where, K(x i ,x j ) represents the kernel function, representing the similarity between sample x i and sample x j ; x i and x j are input vectors;
[0028] S43, test the performance of the support vector machine model by using a test machine, and achieve the required accuracy;
[0029] S44, input the image of the packaging bag to be detected, and perform category judgment to obtain a classification result;
[0030] S5, constructing a template matching model and judging the category; specifically including the following steps:
[0031] S51, constructing a plurality of template images according to the packaging bag to be identified, converting the template images into grayscale images, and obtaining the grayscale values of each region of the template images;
[0032] S52, convert the collected image into a gray scale image, and identify the gray scale image of the collected image and the gray scale image of each template image using a template matching algorithm, set an identification threshold, when the correlation exceeds the threshold, it is judged that the packaging bag to be identified is likely to be the category of the template image, the formula of the template matching algorithm used is as follows:
[0033] Wherein, R(x, y) represents the correlation degree of the region with (x, y) as the center in the to-be-identified image and the template, the greater R(x, y) is, the higher the similarity is, and the higher the possibility of the to-be-identified packaging bag being the template category is, T is the template image, I is the matching image, (x', y') represents all pixel coordinates of the template image, T'(x', y') represents the pixel value of the template image at (x', y') position, I'(x+x', y+y') represents the pixel value of the to-be-identified image at (x+x', y+y') position;
[0034] S53, compare the matching degrees of each template of the to-be-identified packaging bag, and take the category with the highest matching degree as the final classification result;
[0035] S6, construct a neural network model and solve it, specifically including the following steps:
[0036] S61, collect a variety of packaging bag images, take a part of them to construct a sample training set, and take another part to construct a test set;
[0037] S62, initialize the neural network parameters, the BP neural network contains three layers: input layer, hidden layer and output layer; each layer contains a number of neurons, the number of neurons is as follows:
[0038] Wherein, n is the number of input layer neurons, m is the number of output layer neurons, and a is a constant between 1 and 10;
[0039] The weight term update formula is as follows:
[0040] Wherein, W (k+1) represents the K+1th iteration, that is, the updated weight; W (k) represents the Kth iteration, that is, the weight before updating; μ represents the learning rate, which controls the step size of each update; represents the gradient of the weight W, that is, the direction of the weight change; L represents the loss function;
[0041] The bias term update formula is as follows:
[0042] Wherein, b (k+1)b(k+1) represents the bias value after the k+1 iteration, that is, the updated bias value; b (k) b(k) represents the bias value of the k iteration, that is, the bias value before updating; b(k) represents the bias value of the k iteration, that is, the bias value before updating;
[0043] S63, the activation function is selected as the Sigmoid function, and the formula is as follows:
[0044] S64, forward propagation calculation, and the formula of the forward propagation function is as follows: h (l) = f(W (l) h (l-1) +b (l) ) (13)
[0045] where h (l-1) represents the output vector of the previous layer, and also the input vector of the current layer, h (l) represents the output vector of the current layer, f represents the activation function of the current layer, W (l) represents the weight matrix of the previous layer and the current layer, b (l) represents the bias of the current layer;
[0046] S65, error calculation, and the formula of the error loss objective function is as follows: CE(y,y') = -ylog(y')-(1-y)log(1-y') (14)
[0047] where y represents the true label, and y' represents the prediction value of the model;
[0048] S66, the weights and biases in the neural network are updated by the back propagation algorithm to reduce the loss function;
[0049] S67, repeat steps S64-S66, and train the neural network multiple times until convergence or the maximum number of iterations is reached;
[0050] S68, evaluate the performance of the trained neural network model using the test set;
[0051] S69, use the trained neural network model to classify the packaging bags to be detected;
[0052] S7, the results obtained by the three classification methods are subjected to a majority vote mechanism, and the final classification result is obtained;
[0053] S8, classify and stack the packaging bags;
[0054] Further, the classification robot to which the feature fusion classification method is applied comprises a packaging bag, a conveying belt, a camera module, a mechanical arm and a stacking table; the packaging bag is transported to the camera module through the conveying belt, image acquisition is performed on the packaging bag to be detected, then the camera module acquires the image, the classification module is used to classify the image acquired by the camera module, the final category is determined, and finally the mechanical arm places the packaging bag on the stacking table of the corresponding category; the classification module stores the computer program of the feature fusion classification method for a plurality of categories of packaging bags.
[0055] The present application has the following advantages over the prior art:
[0056] (1) The scheme is based on the classification method of the random forest idea of the support vector machine, template matching and neural network, which can improve the packaging bag classification precision and reduce the misrecognition;
[0057] (2) The scheme can improve the automation and intelligent level of the factory and the production efficiency, thereby improving the product quality and market competitiveness of the factory.
[0058] Of course, implementing any product of the present application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS
[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0060] Fig. 1 is a method flowchart of the present application.
[0061] Fig. 2 is a packaging bag recognition schematic diagram of the present application.
[0062] Fig. 3 is a support vector machine method flowchart of the present application.
[0063] Fig. 4 is a template matching method flowchart of the present application.
[0064] Fig. 5 is a bp neural network method flowchart of the present application.
[0065] In the drawings, the components represented by each number are listed as follows: 1 - packaging bag, 2 - conveying belt, 3 - camera module, 4 - mechanical arm, 5 - stacking table. DETAILED DESCRIPTION
[0066] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of the present application.
[0067] As shown in FIG. 2, the classification robot corresponding to the feature fusion classification algorithm for multiple types of packaging bags includes the following parts: 1-packaging bag, 2-conveying belt, 3-camera module, 4-robot arm and 5-palletizing table. First, the packaging bag is transported to the camera module by the conveying belt, and the image of the packaging bag to be detected is collected. Then, the classification module classifies the image by three methods to determine the final category. Finally, the robot arm places the packaging bag on the palletizing table of the corresponding category.
[0068] As shown in FIG. 1, a feature fusion classification algorithm for multiple types of packaging bags includes the following steps:
[0069] S1: The conveying belt transports the packaging bag to the camera module, and the camera module installed on the packaging machine collects the image of the packaging bag to be detected.
[0070] S2: The image denoising module is used to process the collected packaging bag image to improve the image quality and reduce interference.
[0071] S3: The image enhancement module is used to enhance the packaging bag image.
[0072] S4: A support vector machine model is constructed and solved, as shown in FIG. 3, which includes the following steps.
[0073] S41: Collect multiple types of packaging bag images, take a part to construct a sample training set, and take another part to construct a test set.
[0074] S42: A support vector machine model is constructed, and the training set is used to train the support vector machine model. The construction of the support vector machine includes the following steps:
[0075] Establish the objective function of the shortest distance between the feature points and the hyperplane s.t.y i (w T x i +b)≥1,i=1,2…n. (2)
[0076] where w represents an n-dimensional vector, b represents a real number, and s.t. represents a constraint condition. y i represents the label of the i-th sample, and represents 1 or -1, representing positive samples and negative samples, respectively.i is a n-dimensional vector.
[0077] To solve the objective function, a Lagrange function is introduced, and the Lagrange function of the above formula can be written as:
[0078] wherein, α = (α1+α2, …, α n ) Y w represents the normal vector of the partition hyperplane, x i represents the i-th feature attribute of a single sample, and b represents the distance between the origin and the hyperplane.
[0079] The corresponding dual problem is analyzed:
[0080] wherein, α i and α j are Lagrange multipliers, used to solve the dual problem, which is a vector, and each element corresponds to a sample. y i and y j represent the i-th and j-th sample labels, respectively, representing 1 or -1, representing positive samples and negative samples, respectively. x i T is the feature vector of the i-th sample, which is an n-dimensional vector, and T represents transposition. x j is the feature vector of the j-th sample, which is also an n-dimensional vector.
[0081] The optimal solution is obtained
[0082] The optimal classification function is obtained:
[0083] wherein, Sgn() represents a sign function, outputting 1 when the input is greater than 0, outputting -1 when the input is less than 0, and outputting 0 when the input is equal to 0. f(x) represents a classification function, used to calculate which category the sample x belongs to.
[0084] The support vector machine model uses a linear kernel function, and the formula is as follows: K(x i ,x j ) = (x i *x j ) (7)
[0085] wherein, K(x i ,x j ) represents a kernel function, representing the similarity between sample x i and sample x j . x i and x j are input vectors.
[0086] S43: Test the support vector machine model performance with the testing machine to reach the required precision.
[0087] S44: Input the image of the packaging bag to be detected to perform category judgment and obtain a classification result.
[0088] S5: Construct a template matching model and judge the category, as shown in FIG. 4, which specifically includes the following steps:
[0089] S51: Construct multiple template images according to the packaging bag to be identified, convert them into grayscale images, and obtain the grayscale values of each region of the template images.
[0090] S52: Convert the collected image into a grayscale image, and use a template matching algorithm to identify the grayscale image of the collected image and the grayscale image of each template image. Set an identification threshold. When the correlation exceeds the threshold, it is judged that the packaging bag to be identified is likely to be of the category of the template image. The formula of the template matching algorithm used is as follows:
[0091] wherein R(x, y) represents the correlation of the region with the center at (x, y) in the image to be identified with the template. The greater R(x, y) is, the higher the similarity is, and the higher the possibility of the packaging bag to be identified being of the template category is. T is the template image, I is the matching image, (x', y') represents all the pixel coordinates of the template image, T'(x', y') represents the pixel value of the template image at the position (x', y'), and I'(x+x', y+y') represents the pixel value of the image to be identified at the position (x+x', y+y').
[0092] S53: Compare the matching degrees of each template of the packaging bag to be identified, and take the category with the highest matching degree as the final classification result.
[0093] [Corrected according to Rule 91 on 22.10.2024] S6: Construct a neural network model and solve it; as shown in FIG. 5, which specifically includes the following steps:
[0094] S61: Collect multiple category packaging bag images, take a part of them to construct a sample training set, and take another part to construct a test set.
[0095] S62: Initialize the neural network parameters. The BP neural network contains three layers: an input layer, a hidden layer, and an output layer. Each layer contains a number of neurons. The number of neurons is calculated according to the following formula:
[0096] wherein n is the number of input layer neurons, m is the number of output layer neurons, and a is a constant between 1 and 10.
[0097] The weight term update formula is as follows:
[0098] where W (k+1) represents the K+1th iteration, that is, the updated weight. W (k) represents the Kth iteration, that is, the weight before updating. μ represents the learning rate, which controls the step size of each update. represents the gradient of the weight W, that is, the direction of weight change. L represents the loss function.
[0099] The bias term update formula is as follows:
[0100] where b (k+1) represents the bias value after the k+1th iteration, that is, the updated bias value. b (k) represents the bias value of the kth iteration, that is, the bias value before updating. represents the gradient of the bias b, that is, the derivative of the loss function with respect to the bias b,
[0101] S63: Select the Sigmoid function as the activation function, and the formula is as follows:
[0102] S64: Forward propagation calculation, and the forward propagation function formula is as follows: h (l) = f(W (l) h (l-1) + b (l) ) (13)
[0103] where h (l-1) represents the output vector of the previous layer, which is also the input vector of the current layer, h (l) represents the output vector of the current layer, f represents the activation function of the current layer, W (l) represents the weight matrix of the previous layer and the current layer, b (l) represents the bias of the current layer.
[0104] S65: Calculate the error, and the error loss objective function formula is as follows: CE(y,y') = -ylog(y')-(1-y)log(1-y') (14)
[0105] where y represents the true label, and y' represents the predicted value of the model.
[0106] S66: Update the weights and biases in the neural network through the backpropagation algorithm to reduce the loss function.
[0107] S67: Repeat steps S64-S66 to train the neural network multiple times until convergence is reached or the maximum number of iterations is reached.
[0108] S68: Evaluate the performance of the trained neural network model using the test set.
[0109] S69: Classify the to-be-detected packaging bags using the trained neural network model.
[0110] S7: Perform a majority-voting mechanism on the results obtained by the three classification methods to obtain a final classification result.
[0111] S8: Classify and stack the packaging bags.
[0112] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The preferred embodiments do not describe all the details and limit the present application to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of the present application. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and utilize the present application. The present application is limited only by the claims and their full scope and equivalents.
Claims
1. A feature fusion classification method for multi-class packaging bags, characterized in that, It comprises the following steps: S1, the conveyor belt transports the packaging bag to the camera module, and the camera module installed on the packaging machine collects the image of the packaging bag to be detected; S2, the image noise reduction module is used to process the collected packaging bag image, improve the image quality, and reduce the interference; S3, the image enhancement module is used to enhance the packaging bag image; S4, the support vector machine model is constructed and solved; this step comprises the following steps: S41, collect a plurality of types of packaging bag images, take a part to construct a sample training set, and take another part to construct a test set; S42, the support vector machine model is constructed, and the support vector machine model is trained by using the training set; Wherein, the support vector machine model comprises the following steps: Establishing a target function of the shortest distance between a feature point and a hyperplane s.t.y i (w T x i +b)≥1,i=1,2…n. (2) where w represents an n-dimensional vector, b represents a real number, s.t. represents a constraint condition; y i represents the label of the i-th sample, represents 1 or -1, representing positive samples and negative samples, respectively; x i represents the feature vector of the i-th sample, is an n-dimensional vector; To solve the objective function, the Lagrange function is introduced, and the Lagrange function of the above formula is written as: where a = (a1+ a2,..., aM)T, and n ) T ,w represents the normal vector of the dividing hyperplane, x i represents the i-th feature attribute of a single sample, and b represents the distance between the origin and the hyperplane. Analyzing the corresponding dual problem: where α i and α j are Lagrange multipliers, is a vector, each element of which corresponds to a sample, respectively; y i and y j represent the labels of the i-th and j-th samples, representing 1 or -1, representing positive samples and negative samples, respectively; x i T represents the feature vector of the i-th sample, which is an n-dimensional vector, and T represents transposition; x j : the feature vector of the j-th sample, which is also an n-dimensional vector; solving the optimal solution The optimal classification function is obtained: Wherein, Sgn() represents a sign function, when the input is greater than 0, output 1, less than 0, output-1, equal to 0, output 0; f(x) represents a classification function, which is used to calculate which category the sample x belongs to; The support vector machine model adopts a linear kernel function, and the formula is as follows: K(x i ,x j ) = (x i *x j ) (7) where K(x i ,x j ) represents a kernel function, representing the similarity between sample x i and sample x j ; x i and x j are input vectors; S43, the performance of the support vector machine model is tested by using the test machine, and the required precision is reached; S44, the image of the packaging bag to be detected is input, the category is judged, and the classification result is obtained; S5, the template matching model is constructed and the category is judged; specifically comprising the following steps: S51, a plurality of template images are constructed according to the packaging bag to be identified, converted into gray scale images, and the gray scale values of each region of the template image are obtained; S52, convert the collected image into a gray scale image, and use a template matching algorithm to identify the gray scale image of the collected image and the gray scale image of each template image, set an identification threshold, and when the correlation exceeds the threshold, determine that the packaging bag to be identified is likely to be the category of the template image, and the formula of the template matching algorithm used is as follows: Wherein, R(x,y) represents the correlation degree of the region with the center position (x,y) in the image to be identified and the template, the greater R(x,y) is, the higher the similarity is, and the higher the possibility of the packaging bag to be identified belonging to the template category is, T is the template image, I is the matching image, (x',y') represents all pixel coordinates of the template image, T'(x',y') represents the pixel value of the template image at the position (x',y'), and I'(x+x',y+y') represents the pixel value of the to-be-identified image at the position (x+x',y+y'); S53, the matching degrees of each template of the to-be-identified packaging bag are compared, and the category with the highest matching degree is taken as the final classification result; S6, the neural network model is constructed and solved, specifically comprising the following steps: S61, a plurality of types of packaging bag images are collected, a part of which is taken to construct a sample training set, and another part is taken to construct a test set; S62, initialize the neural network parameters, BP neural network contains three layers: input layer, hidden layer and output layer; each layer contains a number of neurons, the number of neurons formula as follows: Wherein, n is the number of input layer neurons, m is the number of output layer neurons, and a is a constant between 1 and 10; The weight term update formula is as follows: where W (k+1) represents the K+1th iteration, i.e., the updated weight; W (k) represents the Kth iteration, i.e., the weight before updating; μ represents the learning rate, which controls the step size of each update; denotes the gradient of the weights W, i.e. the direction of the weight change; L denotes the loss function; The bias term update formula is as follows: wherein b (k+1) represents the bias value after the k+1th iteration, that is, the updated bias value; b (k) represents the bias value of the kth iteration, that is, the bias value before updating; a gradient of the bias b, i.e. a derivative of the loss function with respect to the bias b, S63, the activation function is selected as the sigmoid function, and the formula is as follows: S64, forward propagation calculation, and the forward propagation function formula is as follows: h (l) = f(W (l) h (l-1) + b (l) ) (13) where h (l-1) represents the output vector of the previous layer, which is also the input vector of the current layer, h (l) represents the output vector of the current layer, f represents the activation function of the current layer, W (l) represents the weight matrix of the previous layer and the current layer, b (l) represents the bias of the current layer; S65, error calculation, and the error loss objective function formula is as follows: CE(y,y')=-ylog(y')-(1-y)log(1-y') (14) Wherein, y represents the true label, and y' represents the predicted value of the model; S66, the weights and biases in the neural network are updated by using the back propagation algorithm to reduce the loss function; S67, repeat steps S64-S66, and the neural network is trained multiple times until convergence or The maximum number of iterations is reached; S68, the performance of the trained neural network model is evaluated by using the test set. S69, using the trained neural network model to classify the to-be-detected packaging bag; S7, using the majority rule to vote the results obtained by the three classification methods to obtain the final classification result; S8, classifying and stacking the packaging bags.
2. The feature fusion classification method for multi-class packaging bags according to claim 1, characterized in that, The classification robot applying the feature fusion classification method comprises a packaging bag (1), a conveying belt (2), a camera module (3), a mechanical arm (4), and a stacking table (5). The packaging bag (1) is transported to the camera module (3) through the conveying belt (2), and the image of the to-be-detected packaging bag (1) is collected. Then, the classification module is used to classify the image collected by the camera module (3), determine the final category, and finally the mechanical arm (4) places the packaging bag (1) on the stacking table (5) of the corresponding category. The classification module stores the computer program of the feature fusion classification method for multiple types of packaging bags.