Recycled aggregate crushing value index intelligent detection equipment and method based on multi-scale attention mechanism

Abstract

The invention discloses a recycled aggregate crushing value index intelligent detection device and method based on a multi-scale attention mechanism, and belongs to the technical field of building material intelligent detection.The recycled aggregate crushing value index intelligent detection device comprises a support, a horizontal conveying belt is installed at the top of the support, and inclined conveying belts are installed at the top end and the bottom end of the horizontal conveying belt correspondingly; a baffle is installed at the joint of the horizontal conveying belt and the inclined conveying belt at the bottom end, a vibrator is installed on the horizontal conveying belt, a camera device is installed on the side face of the horizontal conveying belt, and an RA collecting box is installed at the bottom of the inclined conveying belt. According to the intelligent detection equipment and method for the recycled aggregate crushing value index based on the multi-scale attention mechanism, the limitation of a traditional detection method is overcome, and efficient and accurate detection of the RA crushing value index is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of building materials, and in particular to an intelligent detection device and method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism. Background Art

[0002] The crushing value index determines the resistance of recycled aggregates (RA) to crushing, which can directly affect the compressive strength of recycled concrete and is an important basis for classifying its engineering uses. However, in the face of RA with an annual output of up to hundreds of millions of tons, the detection method of its crushing value index highly depends on manual labor, and generally has problems such as high cost, long time consumption, and low sample representativeness, and cannot meet the urgent needs of large-scale, batch, and real-time non-destructive detection and classification of the crushing value index of RA in the process of building automation. Therefore, there is an urgent need to find a method for quickly detecting and classifying the crushing value index of RA to improve the utilization efficiency of large-scale solid waste resources in China. In view of the lack of relevant intelligent detection methods in this regard, an intelligent detection method and device for the crushing value index of recycled aggregates based on a multi-scale attention mechanism are proposed. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent detection device and method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism to overcome the limitations of traditional detection methods and achieve efficient and accurate detection of the crushing value index of RA.

[0004] To achieve the above purpose, the present invention provides an intelligent detection device for the crushing value index of recycled aggregates based on a multi-scale attention mechanism, including a support, a horizontal conveyor belt is installed on the top of the support, inclined conveyor belts are installed at the top and bottom of the horizontal conveyor belt respectively, a baffle is installed at the connection of the horizontal conveyor belt and the inclined conveyor belt at the bottom, a vibrator is installed on the horizontal conveyor belt, a camera device is installed on the side of the horizontal conveyor belt, and an RA collection box is installed at the bottom of the inclined conveyor belt.

[0005] An intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism includes the following steps:

[0006] S1. Collect multiple images containing label information under the same crushing value index;

[0007] S2. Screen and preprocess the RA images;

[0008] S3. Construct a network model CrushNet.

[0009] Preferably, in S1, the crushing value index tests are respectively carried out on different batches and the same batch of RA collected in the RA collection box to obtain the label information of the RA images, and the label information includes the type of RA parent material and grading information.

[0010] Preferably, the screening step of the RA images in S2 is to eliminate the blurred RA images and the RA images with a relatively high content of impurities.

[0011] Preferably, the preprocessing step in S2 is to represent the change in the shooting distance during the image shooting process by the change in the projected area of the standard ball in the RA image, and perform a scaling transformation on the images screened in S2 during the image processing to make the projected area of the sphere in them consistent with the area of the sphere in the standard image. The specific steps are as follows:

[0012] S2.1. Place a standard ball with a diameter D of 2 cm in a batch of RAs taken, then set the standard shooting height h and shoot the RAs and save them as standard images;

[0013] S2.2. Place a standard ball on the surface of the RA during subsequent shooting processes, and its position can be different. After shooting all the images of this batch of RAs, compare the area of the sphere in this batch of images with the area of the sphere in the standard image, and uniformly scale the images with a relatively large deviation in the area of the standard ball to be the same as the area of the sphere in the standard image;

[0014] S2.3. Denote the standard image as image A, the projected area of the sphere in image A as S, denote the image to be transformed as image B, and the projected area of the sphere in image B as S'. Then the scaling ratio α of image B is as shown in formula (1):

[0015]

[0016] S2.4. Perform preprocessing on the transformed RA images, including size adjustment and normalization processing, and crop the size of the RA images to 1000×1000 pixels.

[0017] Preferably, the specific steps for constructing the network model CrushNet in S3 are as follows:

[0018] S3.1. Expand the preprocessed RA images and related labels to form an original image set, and the expansion process is as shown in formulas (2), (3), and (4):

[0019] F1 = G(X) (2);

[0020] F2 = G1(F1)(3);

[0021] F3 = G2(F1)(4);

[0022] Among them, X represents the input image, F1, F2, and F3 represent the output images during the expansion process, and G, G1, and G2 respectively represent the grayscale conversion, rotation transformation, and flipping transformation functions;

[0023] S3.2. Crop the image set obtained from the above operations from 1000×1000 to 224×224 pixel size to reduce the memory of the input model and accelerate the convergence of the model;

[0024] S3.3. Divide the cropped RA images labeled with corresponding crushing value index labels into a training set, a validation set, and a test set according to a set ratio;

[0025] S3.4. Construct a network model CrushNet using channel separation technology. The network model CrushNet includes a channel separation unit, a multi-scale feature extraction unit, a feature dimensionality reduction unit, and a feature fusion unit.

[0026] Preferably, the channel separation technology constructs an attention mechanism by setting a main feature extraction channel and an auxiliary feature extraction channel. The specific steps are as follows:

[0027] S3.4.1. Use a 14×14 convolutional kernel as the main channel, i.e., channel 1, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module after passing through the max pooling layer. Use a 7×7 convolutional kernel as the auxiliary channel, i.e., channel 2, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module after passing through the max pooling layer. After each convolution, the ReLU function is used to alleviate the vanishing gradient problem, as shown in formula (5):

[0028] ReLU = max(0, x) (5);

[0029] where x represents the pixel value of the input feature map;

[0030] S3.4.2. The multi-scale feature extraction module uses the Inception module. The Inception module further extracts the features output from the above two channels from 4 scales and concatenates the extracted features at the output layer as the input to the feature dimensionality reduction unit;

[0031] S3.4.3. Both the main channel and the auxiliary channel correspond to a feature dimensionality reduction unit. The feature dimensionality reduction unit corresponding to the main channel consists of 5 layers of neurons, i.e., Linear(1024, 512) to Linear(64, 1). The feature dimensionality reduction unit corresponding to the auxiliary channel consists of two layers of neurons, i.e., Linear(1024, 512) to Linear(512, 1). The forward propagation and backward propagation of the input features are performed in parallel through the two feature dimensionality reduction units, and finally the reduced features are fused at the output end as the final crushing value index output value.

[0032] Preferably, during the feature dimensionality reduction process, L is used as the loss function value, as shown in formula (6),

[0033]

[0034] Among them, y i is the true value of the i-th image under a certain crushing value index;

[0035] Perform gradient update according to formulas (7) and (8),

[0036]

[0037] Substitute the updated weight w and bias coefficient b into formula (9) to obtain the output of the neuron. After multiple iterations, when the L value reaches the minimum, save the parameters w and b, and substitute them into formula (9) to output the predicted value of the crushing value index,

[0038]

[0039] Among them, L represents the training loss; N represents the total number of samples; k represents the number of neurons; w t represents the weight at the t-th parameter update; b t represents the bias coefficient at the t-th parameter update; η represents the learning rate; represents the gradient of the loss function L(w t ) with respect to the parameter w t ; represents the gradient of the loss function L(b t ) with respect to the parameter b t ; f(w, x, b) i represents the output value of the i-th neuron;

[0040] Record the predicted value of the RA image under the same crushing value index as y i , then the final predicted value of the crushing value index is shown in formula (10):

[0041]

[0042] Among them, n is the number of images, is the predicted value of the i-th image under a certain crushing value index.

[0043] Therefore, the present invention adopts the above-mentioned intelligent detection device and method for the crushing value index of recycled aggregate based on a multi-scale attention mechanism to overcome the limitations of traditional detection methods and achieve efficient and accurate detection of the crushing value index of RA.

[0044] Next, through the accompanying drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0045] Figure 1It is a schematic structural diagram of an intelligent detection device for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention;

[0046] Figure 2 It is a schematic structural diagram of a vibrator of an intelligent detection device for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention;

[0047] Figure 3 It is a schematic diagram showing the change of the image scaling ratio with the shooting distance, a feature difference diagram, and a diagram showing the change of the standard ball area with the scaling ratio of an intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention;

[0048] Figure 4 It is a structural diagram of the network model CrushNet of an intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention;

[0049] Figure 5 It is a structural diagram of the Inception module of an intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention;

[0050] Figure 6 It is the prediction result of the network model CrushNet of an intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism according to the present invention on the crushing value index of RA.

[0051] Reference numerals

[0052] 1, support; 2, horizontal conveyor belt; 3, inclined conveyor belt; 4, baffle; 5, vibrator; 6, imaging device; 7, RA collection box. Detailed implementation manners

[0053] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0054] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before the term cover the elements or objects listed after the term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0055] Example 1

[0056] As Figures 1 to 2 shown, the present invention provides an intelligent detection device for the crushing value index of recycled aggregates based on a multi-scale attention mechanism, including a support 1. A horizontal conveyor belt 2 is installed on the top of the support 1. Oblique conveyor belts 3 are installed at the top and bottom of the horizontal conveyor belt 2 respectively. A baffle 4 is installed at the connection between the horizontal conveyor belt 2 and the bottom oblique conveyor belt 3. A vibrator 5 is installed on the horizontal conveyor belt 2. A camera device 6 is installed on the side of the horizontal conveyor belt 2. An RA collection box 7 is installed at the bottom of the oblique conveyor belt 3.

[0057] When the RA enters the oblique conveyor belt 3, the horizontal conveyor belt 2 is closed. When a batch of RA enters the horizontal conveyor belt 2, the vibrator 5 is turned on (select the vibration power that can make the RA reach the tumbling effect), and at the same time, the wireless camera device 6 is started at the computer end for taking pictures (the specific quantity can be increased according to the actual engineering needs). When the pictures of this batch of RA are taken, the horizontal conveyor belt 2 is started, the baffle 4 is opened, and the RA is conveyed to the RA collection box 7 for the crushing value index test. Repeat the above steps to obtain the images and crushing value indexes of different batches of RA, and import them into the CrushNet model in the computer to complete the training.

[0058] After the training stage is completed, the RA to be tested is collected, and the above steps are repeated. It is conveyed to the horizontal conveyor belt 2 and photographed under different tumbling effects. At the same time, the images are imported into the CrushNet model to obtain the predicted values of the crushing value indexes of the RA with the current image sequence marked. By taking the average of the predicted values of all RA images marked with the same batch, the final crushing value index is obtained.

[0059] An intelligent detection method for the crushing value index of recycled aggregates based on a multi-scale attention mechanism includes the following steps:

[0060] S1. Collect multiple images containing label information under the same crushing value index;

[0061] The crushing value index tests are respectively carried out on different batches and the same batch of RA collected in the RA collection box 7 to obtain the label information of the RA images. The label information includes the types of RA base materials and gradation information, preparing for the training stage of the model.

[0062] S2. Screen and preprocess the RA images;

[0063] The screening steps of the RA images are to eliminate the blurred RA images and the RA images with a high content of sundries, making relevant preparations for the preprocessing and image expansion.

[0064] Considering that in the actual process of taking RA images, different imaging devices, imaging angles, and imaging environments will lead to significant differences in the shooting distances of different RA images. For example, Figure 3 as shown in (a) of Figure 3 , if all images are blindly cropped uniformly and directly imported into model training, the situation shown in (b) of Figure 3 will occur. Although the image sizes are the same, it will result in poor correspondence of the RA feature details contained in the same pixel between the finally input test images and the training images in the model, thereby introducing significant uncertainties in the model training and test results and leading to a large error. Therefore, the change in the projected area of the standard sphere in the RA image is used to represent the change in the shooting distance during the image shooting process. During the image processing, the images screened in S2 are scaled so that the projected area of the sphere in them is the same as the area of the sphere in the standard image. As shown in (c) of Figure 3 , the specific steps are as follows:

[0065] S2.1. Place a standard sphere with a diameter D of 2 cm in a batch of captured RAs, then set the standard shooting height h and capture the RAs and save them as standard images;

[0066] S2.2. Place standard spheres on the surface of the RAs during subsequent shooting processes. Their positions can be different. After capturing all the images of this batch of RAs, compare the sphere areas in this batch of images with the sphere area in the standard image. Uniformly scale the images with a large deviation in the standard sphere area to be the same as the sphere area in the standard image;

[0067] S2.3. Denote the standard image as image A, the projected area of the sphere in image A as S, denote the image to be transformed as image B, and the projected area of the sphere in image B as S'. Then the scaling ratio α of image B is as shown in formula (1):

[0068]

[0069] S2.4. Preprocess the transformed RA images, including size adjustment and normalization. Crop the RA image size to 1000×1000 pixels. In actual applications, a larger image size can be set according to the performance of the device hardware. After normalization processing, convert the image pixel values to between 0 and 1 to accelerate the convergence of the model.

[0070] S3. Build the network model CrushNet.

[0071] The specific steps for building the network model CrushNet are as follows:

[0072] S3.1. Expand the preprocessed RA images and related labels to form an original image set. The expansion process is as shown in formulas (2), (3), and (4):

[0073] F1 = G(X)(2);

[0074] F2 = G1(F1)(3);

[0075] F3 = G2(F1)(4);

[0076] Wherein, X represents the input image, F1, F2, and F3 represent the output images during the expansion process, and G, G1, and G2 respectively represent the grayscale conversion, rotation transformation, and flipping transformation functions;

[0077] S3.2. Crop the image set obtained from the above operations from a size of 1000×1000 to a pixel size of 224×224 to reduce the memory of the input model and accelerate the convergence of the model (it can also be scaled to other sizes);

[0078] S3.3. Divide the cropped RA images labeled with the corresponding crushing value index labels into a training set, a validation set, and a test set according to a set ratio (the number of images in the training set accounts for 85% of the total number);

[0079] S3.4. As Figure 4 shown, use the channel separation technology to construct the network model CrushNet. The network model CrushNet includes a channel separation unit, a multi-scale feature extraction unit, a feature dimensionality reduction unit, and a feature fusion unit.

[0080] The channel separation technology constructs an attention mechanism by setting the main feature extraction channel and the auxiliary feature extraction channel. The specific steps are as follows:

[0081] S3.4.1. Use a 14×14 convolutional kernel as the main channel, that is, channel 1, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module after passing through the max pooling layer; use a 7×7 convolutional kernel as the auxiliary channel, that is, channel 2, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module after passing through the max pooling layer; after each convolution, the ReLU function is passed through to alleviate the gradient disappearance problem, as shown in formula (5):

[0082] ReLU = max(0, x) (5);

[0083] Wherein, x represents the pixel value of the input feature map;

[0084] S3.4.2. As Figure 5As shown, the multi-scale feature extraction module uses the Inception module (the convolution kernel size in this module can adopt the existing parameters of the pre-trained Inception block, or can be further adjusted according to specific image features). The Inception module further extracts the features output from the above two channels from 4 scales respectively, and splices the extracted features at the output layer as the input of the feature dimensionality reduction unit;

[0085] S3.4.3. Both the main channel and the auxiliary channel correspond to a feature dimensionality reduction unit. The feature dimensionality reduction unit corresponding to the main channel consists of 5 layers of neurons, namely Linear(1024, 512) to Linear(64, 1), and the feature dimensionality reduction unit corresponding to the auxiliary channel consists of two layers of neurons, namely Linear(1024, 512) to Linear(512, 1). Forward propagation and backward propagation are performed on the input features in parallel through the two feature dimensionality reduction units, and finally the dimensionality-reduced features are fused at the output end as the output value of the final crushing value index.

[0086] During the feature dimensionality reduction process, L is used as the loss function value as shown in formula (6).

[0087]

[0088] Among them, y i is the true value of the i-th image under a certain crushing value index;

[0089] Gradient update is performed according to formula (7) and formula (8).

[0090]

[0091] Substitute the updated weight w and bias coefficient b into formula (9) to obtain the output of the neuron. After multiple iterations, when the L value reaches the minimum, save the parameters w and b, and substitute them into formula (9) to output the predicted value of the crushing value index.

[0092]

[0093] Among them, L represents the training loss; N represents the total number of samples; k represents the number of neurons; w t represents the weight at the t-th parameter update; b t represents the bias coefficient at the t-th parameter update; η represents the learning rate; represents the gradient of the loss function L(w t ) with respect to the parameter w t ; represents the gradient of the loss function L(b t ) with respect to the parameter b t ; f(w, x, b) iDenote the output value of the \(i\)-th neuron;

[0094] Denote the predicted value of the RA image under the same crushing value index as \(y\) i , then the predicted value of the final crushing value index is shown in formula (10):

[0095]

[0096] where \(n\) is the number of images, is the predicted value of the \(i\)-th image under a certain crushing value index.

[0097] As Figure 6 shown, the predicted value of the crushing value index of RA by the model of the present invention has a high degree of coincidence with the true value, meeting the classification error of the recycled construction waste materials in the current industry standard "Technical Specification for Utilization of Construction Waste in Highway Engineering". In addition, the determination coefficient \(R\) 2 between the predicted value and the true value reaches 98.6%, indicating that the model can accurately predict the change trend of the crushing value index of different RA images, and has the potential to continue training on a large-scale dataset, providing a new idea for the detection of the RA crushing value index.

[0098] To verify the robustness of the present invention, two indicators, namely the average loss and the determination coefficient of the test set, are used for comparative analysis with network structures such as AlexNet, VggNet, and ResNet. The specific data are shown in Table 1.

[0099] Table 1 Prediction results of the network model CrushNet of the present invention and other networks on the test set

[0100] model CrushNet ResNet VggNet AlexNet average loss 0.2％ 0.3％ 0.4％ 0.3％ coefficient of determination 98.6％ 87.2％ 90.5％ 95.1％

[0101] Note: The lower the average loss, the higher the prediction accuracy, and the closer the determination coefficient is to 1, the stronger the model's ability to explain the change of the crushing value index of different images.

[0102] Therefore, the present invention adopts the above-mentioned intelligent detection device and method for the crushing value index of recycled aggregate based on a multi-scale attention mechanism to overcome the limitations of traditional detection methods and achieve efficient and accurate detection of the RA crushing value index.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent detection device for crushing value index of recycled aggregate based on multi-scale attention mechanism, characterized by: It includes a support, a horizontal conveyor belt is installed on the top of the support, oblique conveyor belts are installed on the top and bottom of the horizontal conveyor belt respectively, a baffle is installed at the connection between the horizontal conveyor belt and the oblique conveyor belt at the bottom, a vibrator is installed on the horizontal conveyor belt, a camera device is installed on the side of the horizontal conveyor belt, and an RA collection box is installed at the bottom of the oblique conveyor belt.

2. An intelligent detection method for recycled aggregate crushing value index based on a multi-scale attention mechanism, characterized in that: The following steps are involved: S1, collecting multiple images containing label information under the same crushing value indicator; S2, screening and preprocessing of RA images; S3. Build the network model CrushNet.

3. According to claim 2, the intelligent detection method of recycled aggregate crushing value index based on multi-scale attention mechanism is characterized in that: In S1, the crushing value index test is performed on different batches of RA collected by the RA collection box and the same batch of RA, and the label information of the RA image is obtained. The label information includes the type and gradation information of the RA parent material.

4. According to claim 2, the intelligent detection method of recycled aggregate crushing value index based on multi-scale attention mechanism is characterized in that: The screening step of RA images in S2 is to eliminate blurred RA images and RA images with high impurity content.

5. According to claim 4, the intelligent detection method of recycled aggregate crushing value index based on multi-scale attention mechanism is characterized in that: The preprocessing step in S2 is to use the change of the projection area of ​​the standard sphere in the RA image to represent the change of the shooting distance during the image shooting process. During the image processing process, the image selected in S2 is scaled so that the projection area of ​​the sphere therein is consistent with the area of ​​the sphere in the standard image. The specific steps are as follows: S2.

1. Place a standard ball with a diameter D of 2 cm in a batch of RAs to be photographed, then set a standard photographing height h and photograph the RAs to save as standard images; S2.2, in the subsequent shooting process, a standard ball is placed on the surface of RA, and its position may be different. After all images of the batch of RA are shot, the area of ​​the spheres in the batch of images is compared with the area of ​​the spheres in the standard images. The images with large deviations in the area of ​​the standard spheres are uniformly scaled to the same area as the spheres in the standard images; S2.3, let the standard image be image A, the projection area of ​​the sphere in image A be S, let the image to be transformed be image B, the projection area of ​​the sphere in image B be S', then the scaling ratio α of image B is as shown in formula (1): S2.

4. Preprocess the transformed RA image, including resizing and standardization. The RA image size is cropped to 1000×1000 pixels.

6. The intelligent detection method of recycled aggregate crushing value index based on multi-scale attention mechanism according to claim 2 is characterized in that: The specific steps for building the network model CrushNet in S3 are as follows: S3.

1. Expand the preprocessed RA images and related labels to form the original image set. The expansion process is shown in formulas (2), (3), and (4): F1=G(X) (2); F2=G1(F1) (3); F3=G2(F1) (4); Among them, X represents the input image, F1, F2, and F3 represent the output images in the expansion process, G, G1, and G2 represent the grayscale, rotation transformation, and flip transformation functions respectively; S3.2, crop the image set obtained by the above operation from 1000×1000 to 224×224 pixel size to reduce the memory of the input model and speed up the convergence of the model; S3.3, dividing the cropped RA images labeled with corresponding crushing value index labels into a training set, a validation set, and a test set according to a set ratio; S3.

4. The network model CrushNet is constructed using channel separation technology. The network model CrushNet includes a channel separation unit, a multi-scale feature extraction unit, a feature dimensionality reduction unit and a feature fusion unit.

7. The intelligent detection method for recycled aggregate crushing value index based on multi-scale attention mechanism according to claim 6 is characterized in that: Channel separation technology constructs an attention mechanism by setting the main feature extraction channel and the auxiliary feature extraction channel. The specific steps are as follows: S3.4.

1. A 14×14 convolution kernel is used as the main channel, i.e., channel 1, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module through the maximum pooling layer. A 7×7 convolution kernel is used as the auxiliary channel, i.e., channel 2, to extract features from the original RA image. The extracted features enter the multi-scale feature extraction module through the maximum pooling layer. After each convolution, the ReLU function is used to alleviate the gradient vanishing problem, as shown in formula (5): ReLU = max(0,x) (5); Where x represents the pixel value of the input feature map; S3.4.

2. The multi-scale feature extraction module uses the Inception module. The Inception module further extracts the output features of the above two channels from four scales, and concatenates the extracted features at the output layer as the input of the feature dimension reduction unit. S3.4.

3. Each of the main channel and the auxiliary channel corresponds to a feature dimensionality reduction unit. The feature dimensionality reduction unit corresponding to the main channel consists of five layers of neurons, namely Linear(1024, 512) to Linear(64, 1), and the feature dimensionality reduction unit corresponding to the auxiliary channel consists of two layers of neurons, namely Linear(1024, 512) to Linear(512, 1). The two feature dimensionality reduction units perform forward propagation and backward propagation on the input features in parallel, and finally the reduced features are fused at the output end as the final crushing value indicator output value.

8. The intelligent detection method for crushing value index of recycled aggregate based on multi-scale attention mechanism according to claim 7 is characterized in that: In the process of feature dimensionality reduction, L is used as the loss function value as shown in formula (6): Among them, y i is the true value of the i-th image under a certain crushing value indicator; According to formula (7) and formula (8), the gradient is updated. Substitute the updated weight w and bias coefficient b into formula (9) to get the output of the neuron. After multiple iterations, when the L value reaches the minimum value, save the parameters w and b and substitute them into formula (9) to output the predicted value of the crushing value indicator. Among them, L represents the training loss; N represents the total number of samples; k represents the number of neurons; w t represents the weight of the parameter update for the tth time; b t represents the bias coefficient of the parameter update for the tth time; η represents the learning rate; Denotes the loss function L(w t ) relative to the parameter w t The gradient of Denotes the loss function L(b t ) relative to parameter b t The gradient of f(w,x,b); i Represents the output value of the i-th neuron; The RA image prediction value under the same crushing value index is recorded as y i , then the final predicted value of the crushing value index is shown in formula (10): Where n is the number of images, It is the predicted value of the i-th image under a certain crushing value indicator.

Images