Solid wood panel defect identification method based on deep learning
Through deep learning technology and neural network architecture search algorithm, the subjective error and fatigue problems of manual selection in wood defect detection are solved, and efficient and accurate detection of solid wood board defects is achieved, which is suitable for industrial production and reduces costs.
Patent Information
- Application Number
- CN202411038074.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-12-13
AI Technical Summary
The existing wood defect detection mainly relies on manual selection, which has large subjective errors and fatigue problems, which cannot meet the inspection quality classification needs of industrial production, resulting in high costs and difficult to guarantee.
Using a deep learning-based method, laser profile and color scanners are used to obtain solid wood board information, and a binary classification network is constructed through computer image processing and neural network architecture search algorithms to extract spatial texture features to realize online classification and position recognition of solid wood board defects. .
It improves the efficiency and accuracy of defect detection of solid wood boards, reduces resource waste, ensures the accuracy and speed of inspection, is suitable for industrial production, and reduces costs.
Smart Images
Figure CN119131453A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of solid wood board defect recognition, and specifically to a solid wood board defect recognition method based on deep learning. Background Art
[0002] Wood has a long history of application and a very wide range of applications. Whether it is furniture or books, wood is needed as raw material, so the demand for wood is also huge; but the wood raw materials collected by industry have uneven quality. For example, many wood surfaces have defects such as knots, insect bites and cracks; these defects have a certain impact on the quality of wood, which has caused great obstacles to the processing and treatment of wood in industry, greatly reduced the efficiency of wood processing, and consumed a lot of wood resources; therefore, wood defect detection and classification technology has become an important research topic for scholars, and has received widespread attention and attention, hoping to improve the efficiency of wood processing as much as possible and maximize the value of resources.
[0003] At present, wood defect detection and classification still mainly rely on manual selection. This detection method has disadvantages such as large subjective errors, and human eyes will produce fatigue, so this method cannot meet the needs of industrial production inspection quality grading; coupled with the increasing labor costs, the cost of wood processing is also increasing, and standards cannot be unified, making it difficult to ensure sufficient accuracy in actual detection and classification. Summary of the invention
[0004] Therefore, the present invention solves the technical problem that in the prior art, wood defect detection and classification still mainly rely on manual selection, and this detection method has the disadvantages of large subjective errors, and human eyes will produce fatigue, which cannot meet the needs of industrial production inspection quality grading. At the same time, the labor cost is getting higher and higher, resulting in higher and higher costs for wood processing, and the standards cannot be unified, and it is difficult to ensure sufficient accuracy in actual detection and classification. The present invention provides a solid wood board defect recognition method based on deep learning, which can improve the efficiency and accuracy of solid wood board defect detection.
[0005] The inventive concept of the present invention is: using laser contour and color scanners to obtain the contour and color information of solid wood boards, and transmitting this information to a computer image processing system. The image processing system cuts the solid wood board image into pixel blocks corresponding to the elimination requirements, and extracts spatial texture features to build a machine learning model, performs binary classification training on the presence or absence of defects in the solid wood boards, and then performs further defect feature extraction on the defective solid wood boards, thereby realizing online classification and identification of solid wood board defects.
[0006] The present invention provides a method for identifying defects in solid wood boards based on deep learning, comprising the following steps:
[0007] S1: Collect color images of solid wood panels;
[0008] S2: Obtain pixel blocks using color images;
[0009] S3: Use the neural network architecture search algorithm NAS to build a binary classification network for solid wood boards to achieve the classification of solid wood boards with defects and solid wood boards without defects;
[0010] S4: The confidence optimization strategy is introduced into the binary classification network to ensure that all pixel blocks of solid wood boards with defects are screened out;
[0011] S5: The binary classification network will extract the spatial texture features of the defective solid wood board, and introduce a multi-channel optimization strategy to filter out the more obvious spatial texture features;
[0012] S6: Input the screened spatial texture features and the pixel blocks of the defective solid wood board into the defect detection neural network to further complete the extraction and identification of the defect features of the solid wood board, and realize the identification of the position, shape and type of the defects of the solid wood board.
[0013] Preferably, in S3, the construction of a binary classification network for solid wood boards includes the following steps:
[0014] S31: Determine the search space of the neural network architecture search algorithm based on the task attributes of judging whether the solid wood board has defects;
[0015] S32: Further determine the search strategy of the neural network architecture search algorithm using the following formula:
[0016]
[0017] In the formula, is a network with architecture, N A is the set of all possible network architectures, C valid is the evaluation criterion on the validation dataset, w * Represents the optimal weight value of the network, L train is the loss function on the training data set; st refers to subject to (restricted to) argmin() function, which represents the parameter value at which a function takes the minimum value in its domain.
[0018] S33: Comparing the accuracy and speed of the binary classification network to distinguish between defective solid wood board pixel blocks is used as a performance evaluation strategy for the binary classification network. Considering the detection accuracy and detection speed of the binary classification network, the fitness value F is designed to take these two influencing factors into account. F is expressed as:
[0019] F=2-F s -F acc
[0020] Among them, F s Represents the fitness value of the model detection speed, F acc Represents the average detection accuracy of the model;
[0021]
[0022] Among them, f, f max and f min They represent the number of floating-point operations FLOP of the model and the maximum and minimum floating-point operations of all possible neural architectures, respectively. α and β are parameters used to scale the floating-point operation value.
[0023] Preferably, the calculation formula of the floating-point operation value in S33 is as follows:
[0024]
[0025] Among them, FLOPs CONV is the floating point operation value of the convolution layer, FLOPs maxp is the floating point value of the maximum pooling layer; H, W and C in are the height, width and number of input channels of a feature image; K represents the kernel size of each network layer; C out Indicates the number of output channels;
[0026] The maximum number of network layers in the architecture space is n, and each convolutional layer has the same padding, obtaining the following expression:
[0027]
[0028] Among them, H, W and d are the height, width and depth of the input image respectively; k represents the kernel size of each layer; c represents the input channel of the feature;
[0029] Another factor in the fitness value is the average accuracy of the model, which is defined as:
[0030]
[0031] in, is the number of correct classifications of the categories, C i is the total number of categories, C n Represents the total number of categories.
[0032] Preferably, a confidence optimization strategy is introduced in S4:
[0033]
[0034] Among them, y1 and y2 are the outputs for each category, α is the confidence of the user's choice, and r is the result.
[0035] Preferably, in S1, a laser profile and color scanner is used to collect a color image of the solid wood board.
[0036] Preferably, the color image in S2 is segmented into pixel blocks by a 20*20 pixel size method.
[0037] Preferably, in S1, after the color image of the solid wood board is collected, the color image is corrected.
[0038] Compared with the prior art, the present invention has the following advantages:
[0039] 1. The solid wood board defect recognition method based on deep learning provided by the present invention introduces a binary classification network at the front end of the traditional defect detection algorithm. The network is used to complete the classification of defective wood and non-defective wood in advance, which effectively reduces the waste of defect detection algorithm resources and improves the efficiency of solid wood board defect recognition.
[0040] 2. The solid wood board defect recognition method based on deep learning provided by the present invention utilizes a neural network architecture search algorithm and a genetic algorithm, determines the search space and evaluation strategy of the network architecture search algorithm according to the actual needs of the binary classification network, and realizes the optimization of the binary classification algorithm for solid wood boards, thereby ensuring both the accuracy of the two types of wood classification and the recognition speed of the binary classification algorithm.
[0041] 3. The solid wood board defect recognition method based on deep learning provided by the present invention introduces a confidence optimization strategy after the binary classification algorithm network architecture is built, which can more accurately detect defective solid wood boards at the cost of a slightly higher false alarm rate, thereby improving the stability of the method in industrial production.
[0042] 4. The solid wood board defect recognition method based on deep learning provided by the present invention introduces a multi-channel selection optimization strategy to perform lightweight optimization on redundant multi-channel image features. The optimized model has a significant improvement in detection speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0044] Figure 1 A network structure diagram of a method for identifying defects in solid wood panels based on deep learning provided by the present invention;
[0045] Figure 2 A search curve for the binary classification network architecture in the present invention;
[0046] Figure 3 It is a network structure diagram of the binary classification network in the present invention;
[0047] Figure 4 This is a validation diagram of the effect obtained after training the model of the present invention. DETAILED DESCRIPTION
[0048] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0049] In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limitations on the present invention.
[0050] Example 1
[0051] Figure 1 The network structure diagram of the method for identifying defects in solid wood panels based on deep learning provided in this embodiment is as follows. The target detection algorithm mainly includes three parts:
[0052] (1) Preliminary classification: Each input image is classified through the glance network to determine whether the image is a wood image with defects. If the image is a wood image with defects, it will be output to the subsequent network for the next step of processing.
[0053] (2) Feature extraction: When the glance network performs a preliminary classification on an image, it also extracts the features of the image. These features will be input into the Resnet50 network and feature pyramid network together with the input image to further extract more features.
[0054] (3) Defect Detection: The features extracted by the above networks are input into the RPN network and ROI Align respectively. The RPN network marks the rectangular candidate regions containing defects in the image. RoI Align derives higher-level feature maps from RoI by combining the extracted feature maps and RPN marked regions. These high-level feature maps are then input into several fully connected layers (FC) to obtain the type, coordinates, and mask of the defect.
[0055] In this embodiment, a method for identifying defects in solid wood boards based on deep learning specifically includes the following steps:
[0056] S1: Collect color images of solid wood panels;
[0057] S2: Obtain pixel blocks using color images;
[0058] S3: Use the neural network architecture search algorithm NAS to build a binary classification network for solid wood boards to achieve the classification of solid wood boards with defects and solid wood boards without defects;
[0059] S4: The confidence optimization strategy is introduced into the binary classification network to ensure that all pixel blocks of solid wood boards with defects are screened out;
[0060] S5: The binary classification network will extract the spatial texture features of the defective solid wood board, and introduce a multi-channel optimization strategy to filter out the more obvious spatial texture features;
[0061] S6: Input the screened spatial texture features and the pixel blocks of the defective solid wood board into the defect detection neural network to further complete the extraction and identification of the defect features of the solid wood board, and realize the identification of the position, shape and type of the defects of the solid wood board.
[0062] When the binary classification network extracts the spatial texture features of defective solid wood boards, the features of different channels are similar, which makes some features ineffective for improving the accuracy of subsequent defect detection. However, the input of these features will increase the detection time of the subsequent network and affect the overall efficiency. Therefore, by introducing a multi-channel selection optimization strategy, these redundant channels are screened out in advance to avoid input into the subsequent detection network, thereby improving the speed of defect detection.
[0063] like Figure 2 As shown in the figure, in the initial generation of training, there are various evaluation values, which represent the structural changes of the generated network. As the number of training generations increases, the network points slowly gather together, and some of them behave irregularly due to mutations. Finally, all networks tend to have similar fitness values, which means that the algorithm has reached convergence. The fitness value converges to a point that represents the fitness value of the Glance Network.
[0064] In this embodiment, in S3, the construction of a binary classification network for solid wood boards includes the following steps:
[0065] S31: Combined with the task attributes of judging whether the solid wood board has defects, the search space of the neural network architecture search algorithm is determined; the encoding method of the search space is shown in the following table:
[0066]
[0067] S32: After determining the search space of the neural network architecture search, the network is restricted to a limited depth, but there are still many candidate networks. In order to effectively search for the desired binary classification network in the huge search space, the search strategy of the neural network architecture search algorithm is further determined to avoid premature convergence to the suboptimal architecture area. Here, the search strategy is transformed into a two-level optimization problem;
[0068]
[0069] In the formula, is a network with architecture, N A is the set of all possible network architectures, C valid is the evaluation criterion on the validation dataset, w * Represents the optimal weight value of the network, L train is the loss function on the training dataset;
[0070] In order to solve the two-level optimization problem, a genetic algorithm is used to valid Optimize and use stochastic gradient descent momentum SGD-M to optimize w * optimization.
[0071] S33: In order to achieve a high prediction performance architecture on the data set, the accuracy and speed of the binary classification network in distinguishing between defective solid wood board pixel blocks are compared as a performance evaluation strategy for the binary classification network. Considering the detection accuracy and detection speed of the binary classification network, the fitness value F is designed to take both of these influencing factors into account. F is expressed as:
[0072] F=2-F s -F acc
[0073] Among them, F s Represents the fitness value of the model detection speed, F acc Represents the average detection accuracy of the model;
[0074]
[0075] Among them, f, f max and f minThey represent the number of floating-point operations FLOP of the model and the maximum and minimum floating-point operations of all possible neural architectures, respectively. α and β are parameters used to scale the floating-point operation value.
[0076] like Figure 3 As shown in the figure, after 66 hours of neural network structure search, the network structure of Glance Network was determined. The structure consists of convolutional layers, maximum pooling layers, expansion layers, and fully connected layers.
[0077] In this embodiment, the number of floating-point operations is used as a measure of the speed of the neural network, which is an indirect method to measure the complexity and running speed of the model. The calculation formula of the floating-point operation value in S33 is as follows:
[0078]
[0079] Among them, FLOPs CONV is the floating point operation value of the convolution layer, FLOPs maxp is the floating point value of the maximum pooling layer; H, W and C in are the height, width and number of input channels of a feature image; K represents the kernel size of each network layer; C out Indicates the number of output channels;
[0080] The maximum number of network layers in the architecture space is n, and each convolutional layer has the same padding, obtaining the following expression:
[0081]
[0082] Among them, H, W and d are the height, width and depth of the input image respectively; k represents the kernel size of each layer; c represents the input channel of the feature;
[0083] Another factor in the fitness value is the average accuracy of the model, which is defined as:
[0084]
[0085] in, is the number of correct classifications of the categories, C i is the total number of categories, C n Represents the total number of categories.
[0086] In this embodiment, the confidence optimization strategy is introduced in S4:
[0087]
[0088] Among them, y1 and y2 are the outputs of each category, α is the confidence selected by the user, and r is the result. High confidence makes the model more likely to classify the input as category 1, while negative confidence makes the model tend to classify the input as category 0. The two-classification network with the confidence optimization strategy can more accurately detect defective solid wood boards at the cost of a slightly higher false alarm rate.
[0089] In this embodiment, the laser profile and color scanner are used to collect the color image of the solid wood board in S1.
[0090] In this embodiment, the color image in S2 is segmented into pixel blocks by a 20*20 pixel size method.
[0091] In this embodiment, in S1, after the color image of the solid wood board is collected, the color image is corrected.
[0092] like Figure 4 As shown in the first column, the confidence of live joints is 86.3% and that of dead joints is 96.2%. In the remaining images, the confidence of dead joints, live joints and cracks are 99.8%, 99.2% and 99.7% respectively. Based on these results, we conclude that the detection performance of this method is satisfactory.
[0093] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.
Claims
1. A method for identifying defects in solid wood panels based on deep learning, characterized in that: The steps include: S1: Collect color images of solid wood panels; S2: Obtain pixel blocks using color images; S3: Use the neural network architecture search algorithm to build a binary classification network for solid wood boards to achieve the classification of solid wood boards with defects and solid wood boards without defects; S4: A confidence optimization strategy is introduced into the binary classification network to filter out all pixel blocks of solid wood boards with defects; S5: The binary classification network will extract the spatial texture features of the defective solid wood board, and introduce a multi-channel optimization strategy to filter out the spatial texture features; S6: Input the screened spatial texture features and the pixel blocks of the defective solid wood board into the defect detection neural network to complete the extraction and recognition of the defect features of the solid wood board, and realize the recognition of the position, shape and type of the defects of the solid wood board.
2. The method for identifying defects in solid wood panels based on deep learning according to claim 1, characterized in that: In S3, the construction of the binary classification network for solid wood boards includes the following steps: S31: Determine the search space of the neural network architecture search algorithm based on the task attributes of judging whether the solid wood board has defects; S32: Determine the search strategy of the neural network architecture search algorithm using the following formula: In the formula, is a network with architecture, N A is the set of all network architectures, C valid is the evaluation criterion on the validation dataset, w * Represents the optimal weight value of the network, L train is the loss function on the training dataset; S33: Comparing the accuracy and speed of the binary classification network to distinguish between defective and non-defective solid wood board pixel blocks is used as a performance evaluation strategy for the binary classification network. These two influencing factors are taken into account by designing the fitness value F, which is expressed as: F=2-F s -F acc Among them, F s Represents the fitness value of the model detection speed, F acc Represents the average detection accuracy of the model; Among them, f, f max and f min Represent the number of floating-point operations of the model and the maximum and minimum values of floating-point operations for all possible neural architectures, respectively. α and β are parameters used to scale the floating-point operation value.
3. The method for identifying defects in solid wood panels based on deep learning according to claim 2, characterized in that: The calculation formula of the floating point operation value in S33 is as follows: Among them, FLOPs CONV is the floating point operation value of the convolution layer, FLOPs maxp is the floating point value of the maximum pooling layer; H, W and C in are the height, width and number of input channels of a feature image; K represents the kernel size of each network layer; C out Indicates the number of output channels; The maximum number of network layers in the architecture space is n, and each convolutional layer has the same padding, obtaining the following expression: Among them, H, W and d are the height, width and depth of the input image respectively; k represents the kernel size of each layer; c represents the input channel of the feature; Another factor in the fitness value is the average accuracy of the model, which is defined as: in, is the number of correct classifications of the categories, C i is the total number of categories, C n Represents the total number of categories.
4. The method for identifying defects in solid wood panels based on deep learning according to claim 1, characterized in that: The confidence optimization strategy is introduced in S4: Among them, y1 and y2 are the outputs for each category, α is the confidence of the user's choice, and r is the result.
5. The method for identifying defects in solid wood panels based on deep learning according to claim 1, characterized in that: In S1, a laser profile and color scanner is used to collect a color image of the solid wood board.
6. The method for identifying defects in solid wood panels based on deep learning according to claim 1, characterized in that: The color image in S2 is segmented into pixel blocks with a size of 20*20 pixels.
7. The method for identifying defects in solid wood panels based on deep learning according to claim 1, characterized in that: In S1, after the color image of the solid wood board is collected, the color image is corrected.