Concrete slump detection method based on salient object detection

Abstract

The application discloses a concrete slump detection method based on salient target detection, and specifically comprises the following steps: step 1, image preprocessing; step 2, multi-scale reuse salient target identification based on deep learning; and step 3, pixel difference conversion and result analysis output. By using the application, the problems that the existing detection methods rely on manual operation and subjective judgment are solved, the detection result is more accurate and stable, human resources are saved, the method system of the computer vision and artificial intelligence technical field is enriched, the development of the salient target detection field is supported, and a selection is provided for the concrete slump detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, specifically relating to a method for detecting concrete slump based on salient target detection. Background Technology

[0002] Concrete slump testing is an indicator for assessing the fluidity and plasticity of concrete; it refers to the degree to which concrete collapses in a slump cone test, and is also known as slump value or slump.

[0003] In recent years, concrete slump testing has become increasingly important in engineering construction and quality control. Traditional concrete slump testing methods include slump cone test, pressure plate method, and flow meter. These methods usually rely on manual operation and subjective judgment, which are subject to subjectivity, instability and human resource limitations. Summary of the Invention

[0004] The purpose of this invention is to provide a concrete slump detection method based on salient target detection, which solves the problem that existing detection methods rely on manual operation and subjective judgment.

[0005] The technical solution adopted in this invention is: a concrete slump detection method based on salient target detection, which is implemented according to the following steps:

[0006] Step 1, Image preprocessing;

[0007] Step 2: Multi-scale reuse of salient target identification based on deep learning;

[0008] Step 3: Pixel difference conversion and result analysis output.

[0009] The invention is further characterized by:

[0010] Step 1, image preprocessing, is implemented as follows:

[0011] Step 1.1: Collect concrete sample images with different slump and label them, marking the collapsed areas as salient targets to construct a labeled training set;

[0012] Step 1.2: Process the images in the training dataset to generate similar but not identical samples for data augmentation.

[0013] Step 1.3: Normalize the pixel values ​​of the image to a specific range;

[0014] Step 2, which is based on deep learning for multi-scale reuse of salient target identification, is implemented as follows:

[0015] Step 2.1: Construct a 3×3 standard convolution for feature extraction and adjust the number of channels to half;

[0016] Step 2.2: Construct a multi-scale feature extraction module with ResNet50 as the backbone network to extract multi-scale image features;

[0017] Step 2.3: Construct a cross-bidirectional dense feature aggregation module to perform full-resolution fusion of the features extracted by the backbone network, and extract multi-scale features with richer semantics;

[0018] Step 2.4: Construct a multi-resolution semantic complementarity module to retain useful information in adjacent feature nodes of multi-scale features at multiple levels, and restore the semantic and spatial information of salient targets step by step.

[0019] Step 2.5: Set network parameters and perform network training and testing;

[0020] Step 2.2 is implemented as follows:

[0021] Step 2.2.1: Construct a network with ResNet50 as the backbone;

[0022] Step 2.2.2: Multi-scale feature extraction is achieved by stacking multiple residual blocks;

[0023] Step 2.2.3: After the last residual block, add a global average pooling layer to perform average pooling on the feature map of each channel to obtain a feature vector of fixed length.

[0024] Step 2.3 is implemented in the following steps:

[0025] Step 2.3.1: Input the basic features a1 to a5 extracted from the five stages of the ResNet50 network as input to the first bidirectional dense feature aggregation module, which consists of three bidirectional dense feature aggregation layers.

[0026] Step 2.3.2, the first bidirectional dense feature aggregation layer takes five signals as input and outputs: four signals as input to the next layer and the intermediate multi-scale semantic feature n5;

[0027] Step 2.3.3, the second bidirectional dense feature aggregation layer takes four signals as input and outputs: the three signals as input to the next layer and the intermediate multi-scale semantic feature n4;

[0028] Step 2.3.4, the third bidirectional dense feature aggregation layer, takes three signals as input and outputs three intermediate multi-scale semantic features n1, n2, and n3;

[0029] Step 2.3.5: Input the five intermediate multi-scale semantic features n1, n2, n3, n4, and n5 into the next bidirectional dense feature aggregation module. The two bidirectional dense feature aggregation modules together constitute a cross-bidirectional dense feature aggregation module with the same structure. The final output is the final multi-scale semantic features c1, c2, c3, c4, and c5.

[0030] Step 2.3.2 is implemented in the following steps:

[0031] Step 2.3.2.1: a5 is upsampled and subjected to element-wise linear weighting and convolution with a4 to obtain m4. Similarly, m3, m2 and b1 are obtained from bottom to top.

[0032] Step 2.3.2.2: b1 is downsampled and subjected to element-wise linear weighting and convolution with m2 and a2 to obtain b2. Similarly, b3, b4, and b5 are obtained sequentially from top to bottom. Among them, b5 serves as one of the final multi-scale semantic features c5. Meanwhile, b1, b2, b3, and b4 serve as the input to the next bidirectional dense feature aggregation layer. The fusion formula is as follows:

[0033]

[0034]

[0035] In the formula, conv represents the convolution operation on the features, consisting of convolution, batch normalization, and the ReLU activation function; w i The weight coefficients assigned to each input during the feature fusion stage are initialized to random numbers in the range of (0,1).

[0036] Step 2.4, constructing the multi-resolution semantic complementarity module, is implemented according to the following steps:

[0037] In steps 2.4.1 and 2.3, the final multi-scale semantic features c1, c2, c3, c4, and c5 output by the bidirectional dense feature aggregation module are used as inputs to the multi-resolution semantic complementarity module. The resolution of these features is halved in turn, and they contain semantic information from targets at different scales.

[0038] Step 2.4.2 involves concatting two adjacent input features, followed by dilated convolution with a rate of 2, normalization, and ReLU operations to obtain the global semantic information that fuses the two input features; the formula is as follows:

[0039] f g =conv(concat(f1,f2)) (3)

[0040] In the formula, f1 and f2 represent adjacent features, and conv represents a hole with a batch normalization layer and a ReLU activation function;

[0041] Step 2.4.3 involves adding the obtained global semantic information back into the input features through element-wise summation, thereby enhancing the input features both spatially and semantically; the formula is as follows:

[0042]

[0043]

[0044] In the formula, This is element-wise addition;

[0045] Step 2.4.4 involves selectively inheriting two sets of spatially and semantically reinforced features through element-wise multiplication. This achieves the goal of allowing the input features to inherit important characteristics while discarding more noise. The formula is as follows:

[0046]

[0047] In the formula, This is element-wise multiplication;

[0048] Step 3 is implemented in the following steps:

[0049] Step 3.1: Measure the actual height of the slump bucket. Send the images taken before and after the slump experiment to the multi-scale reuse salient target recognition module trained in Step 2, and output the processed binary images P and P′. The regions with a gray value of 255 are the salient target recognition results for the slump bucket and the cement, respectively.

[0050] Step 3.2: Search the binarized image P obtained in Step 3.1 before the slump test to find the highest point Y1 and the lowest point Y2 in the region with a gray value of 255. h Calculate the absolute difference between the two points on the y-axis coordinates, and then calculate the ratio K of the image size to the actual size based on the actual height of the slump bucket.

[0051] Step 3.3: Search the binarized image P′ obtained after the collapse experiment in Step 3.1 to find the minimum y-axis coordinate value of the region with a gray value of 255, Y1′, calculate the absolute difference between this point and the minimum y-axis coordinate value of the region with a gray value of 255 in P, and then calculate the actual collapse height T according to K.

[0052] Step 3.4: Calculate the concrete classification results based on slump grading. The slump is 10-40mm, which is low plasticity concrete; 50-90mm, which is plastic concrete; 100-150mm, which is fluid concrete; and slump greater than 160mm, which is highly fluid concrete.

[0053] Step 3.5: Output P, ​​P′, T and the concrete classification results.

[0054] The beneficial effects of this invention are:

[0055] In this invention, the concrete slump detection method based on salient target detection does not involve data preprocessing, pixel difference conversion, or result analysis output in the model training process. It utilizes deep learning-based multi-scale reuse of salient target recognition, extracting multi-scale features through a ResNet50 backbone network to diversify the initial features. Furthermore, it incorporates a bidirectional dense feature aggregation module and a multi-resolution semantic complementarity module to achieve multi-layer semantic feature fusion at different resolutions, preventing the dilution of useful features and capturing more discriminative abstract features. Overall, this method first preprocesses the image, then performs salient target recognition to annotate the concrete, and finally completes slump detection in the pixel difference conversion and result analysis output module. Experimental results show that, compared with existing methods, the concrete slump detection method based on salient target detection overcomes the problems of reliance on manual operation and subjective judgment, enriches the methodological system of artificial intelligence and computer vision, and provides an alternative for concrete slump detection. Attached Figure Description

[0056] Figure 1 This is a network architecture diagram of the concrete slump detection method based on salient target detection in step 2 of the present invention;

[0057] Figure 2 This is a network architecture diagram of the bidirectional dense feature aggregation layer in step 2.3 of the concrete slump detection method based on salient target detection in this invention;

[0058] Figure 3 This is a network architecture diagram of the multi-resolution semantic complementarity module in step 2.4 of the concrete slump detection method based on salient target detection of the present invention. Detailed Implementation

[0059] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0060] This invention provides a concrete slump detection method based on salient target detection. With the development of machine learning and deep learning, automatic concrete slump detection has been extensively studied. These methods can identify and measure concrete slump based on features such as color, texture, and shape. Compared with traditional methods, deep learning-based methods can learn feature representations from large amounts of data and automatically extract features useful for slump prediction. The models can adapt to different concrete properties and construction conditions and have high prediction accuracy and stability. This application chooses to use convolutional neural networks as the basis and focuses on salient target detection technology. The key technologies include the following three aspects: image preprocessing, salient target recognition, and pixel difference conversion and result analysis output. Through the implementation of these three key technologies, the task of concrete slump detection is completed, thereby overcoming the problems of existing concrete slump detection methods relying on manual operation and subjective judgment. This is of great significance for realizing technologies such as salient target recognition, slump detection, and image processing.

[0061] Example 1

[0062] The concrete slump detection method based on salient target detection is implemented according to the following steps:

[0063] Step 1, Image preprocessing;

[0064] Step 2: Multi-scale reuse of salient target identification based on deep learning;

[0065] Step 3: Pixel difference conversion and result analysis output.

[0066] Example 2

[0067] The concrete slump detection method based on salient target detection is implemented according to the following steps:

[0068] Step 1, Image Preprocessing:

[0069] Step 1.1: Collect concrete sample images with different slump and label them, marking the collapsed areas as salient targets to construct a labeled training set;

[0070] Step 1.2: Perform random horizontal or vertical rotations on the images in the training dataset to generate similar but not identical samples for data augmentation.

[0071] Step 1.3: Normalize the pixel values ​​of the image to a specific range, such as [0,1] or [-1,1]. Normalization helps to make the image data have similar scales, which is beneficial for model training and convergence.

[0072] Step 2: Multi-scale reuse of salient target recognition based on deep learning, its network architecture Figure 1 As shown;

[0073] Step 2.1: Construct a 3×3 standard convolution for feature extraction and adjust the number of channels to half. The purpose of adjusting the number of channels is to reduce some of the computational burden.

[0074] Step 2.2: Construct a multi-scale feature extraction module with ResNet50 as the backbone network to extract multi-scale image features;

[0075] Step 2.2.1: Construct a network with ResNet50 as the backbone to diversify the initial features;

[0076] Step 2.2.2: Multi-scale feature extraction is achieved by stacking multiple residual blocks. Different levels of these residual blocks provide feature representation capabilities at different scales. In deep residual blocks, the network stride is smaller and the feature map size is smaller, providing higher-level semantic information; in shallow residual blocks, the network stride is larger and the feature map size is larger, providing more detailed information.

[0077] Step 2.2.3: After the last residual block, a global average pooling layer is added to perform average pooling on the feature map of each channel, thereby obtaining a feature vector of fixed length.

[0078] Step 2.3: Construct a cross-bidirectional dense feature aggregation module to perform full-resolution fusion of the features extracted by the backbone network, and extract multi-scale features with richer semantics;

[0079] Step 2.3.1: Input the basic features a1 to a5 extracted from the five stages of the ResNet50 network as input to the first bidirectional dense feature aggregation module. This module consists of three bidirectional dense feature aggregation layers, as shown in Figure 2.

[0080] Step 2.3.2: The first bidirectional dense feature aggregation layer takes five signals as input and outputs: four signals as input to the next layer and the intermediate multi-scale semantic features n5.

[0081] Step 2.3.3, the second bidirectional dense feature aggregation layer, takes four signals as input and outputs: the three signals that will be used as input to the next layer and the intermediate multi-scale semantic features n4.

[0082] Step 2.3.4, the third bidirectional dense feature aggregation layer, takes three signals as input and outputs three intermediate multi-scale semantic features n1, n2, and n3;

[0083] Step 2.3.5: Input the five intermediate multi-scale semantic features n1, n2, n3, n4, and n5 into the next bidirectional dense feature aggregation module. The two bidirectional dense feature aggregation modules together form a cross-bidirectional dense feature aggregation module with identical structures. The final output is the final multi-scale semantic features c1, c2, c3, c4, and c5.

[0084] Step 2.4: Construct a multi-resolution semantic complementarity module to retain useful information from adjacent feature nodes at multiple levels of multi-scale features, and progressively restore the semantic and spatial information of the salient target. Its structure is as follows: Figure 3 As shown;

[0085] Step 2.4.1: The final multi-scale semantic features c1, c2, c3, c4, and c5 output by the bidirectional dense feature aggregation module in the previous step are used as input to the multi-resolution semantic complementarity module. The resolution of these features is halved in turn, and they contain semantic information from targets at different scales.

[0086] Step 2.4.2 involves concatting two adjacent input features, followed by dilated convolution with a rate of 2, normalization, and ReLU operations to obtain the global semantic information that fuses the two input features. The formula is as follows:

[0087] f g =conv(concat(f1,f2)) (3)

[0088] Where f1 and f2 represent adjacent features, and conv represents a hole with a batch normalization layer and a ReLU activation function;

[0089] Step 2.4.3 involves adding the obtained global semantic information back into the input features through element-wise summation, thereby enhancing the input features both spatially and semantically. The formula is as follows:

[0090]

[0091]

[0092] in, This is element-wise addition;

[0093] Step 2.4.4 involves selectively inheriting two sets of spatially and semantically reinforced features through element-wise multiplication. This achieves the goal of allowing the input features to inherit important characteristics while discarding more noise. The formula is as follows:

[0094]

[0095] in, This is element-wise multiplication;

[0096] Step 2.5: Set network parameters and perform network training and testing;

[0097] Step 3: Pixel difference conversion and result analysis output;

[0098] Step 3.1: Measure the actual height of the slump bucket. Send the images taken before and after the slump experiment to the multi-scale reuse salient target recognition module trained in Step 2, and output the processed binary images P and P′. The regions with a gray value of 255 are the salient target recognition results for the slump bucket and the cement, respectively.

[0099] Step 3.2: Search the binarized image P obtained in Step 3.1 before the slump test to find the highest point Y1 and the lowest point Y2 in the region with a gray value of 255. h Calculate the absolute difference between the two points on the y-axis coordinates, and then calculate the ratio K of the image size to the actual size based on the actual height of the slump bucket.

[0100] Step 3.3: Search the binarized image P′ obtained after the collapse experiment in Step 3.1 to find the minimum y-axis coordinate value of the region with a gray value of 255, Y1′, calculate the absolute difference between this point and the minimum y-axis coordinate value of the region with a gray value of 255 in P, and then calculate the actual collapse height T according to K.

[0101] Step 3.4: Calculate the concrete classification results based on slump grading. The slump is 10-40mm, which is low plasticity concrete; 50-90mm, which is plastic concrete; 100-150mm, which is fluid concrete; and slump greater than 160mm, which is highly fluid concrete.

[0102] Step 3.5: Output P, ​​P′, T and the concrete classification results.

[0103] Example 3

[0104] This embodiment takes the bidirectional dense feature aggregation layer as an example, specifically step 2.3.2, and the specific operation is as follows:

[0105] Step 2.3.2.1: a5 is upsampled and subjected to element-wise linear weighting and convolution with a4 to obtain m4. Similarly, m3, m2 and b1 are obtained from bottom to top.

[0106] Step 2.3.2.2: b1 is downsampled and subjected to element-wise linear weighting and convolution with m2 and a2 to obtain b2. Similarly, b3, b4, and b5 are obtained sequentially from top to bottom. b5 serves as one of the final multi-scale semantic features, c5. Meanwhile, b1, b2, b3, and b4 serve as the input to the next bidirectional dense feature aggregation layer. The fusion formula is as follows:

[0107]

[0108]

[0109] In the formula, conv represents the convolution operation on the features, consisting of convolution, batch normalization, and the ReLU activation function; w i The weight coefficients assigned to each input during the feature fusion stage are initialized to random numbers in the range (0,1).

Claims

1. A method for detecting concrete slump based on salient target detection, characterized in that, The specific steps are as follows: Step 1, Image preprocessing; Step 2: Multi-scale reuse of salient target identification based on deep learning; Step 3: Pixel difference conversion and result analysis output; Step 2, the multi-scale reuse of salient target identification based on deep learning, is implemented according to the following steps: Step 2.1: Construct a 3×3 standard convolution for feature extraction and adjust the number of channels to half; Step 2.2: Construct a multi-scale feature extraction module with ResNet50 as the backbone network to extract multi-scale image features; Step 2.3: Construct a cross-bidirectional dense feature aggregation module to perform full-resolution fusion of the features extracted by the backbone network, and extract multi-scale features with richer semantics; Step 2.4: Construct a multi-resolution semantic complementarity module to retain useful information in adjacent feature nodes of multi-scale features at multiple levels, and restore the semantic and spatial information of salient targets step by step. Step 2.5: Set network parameters and perform network training and testing; Step 3 is implemented in the following steps: Step 3.1: Measure the actual height of the slump cone. Input the images taken before and after the slump experiment into the multi-scale reuse salient target recognition module trained in Step 2, and output the processed binarized images. , The regions with a grayscale value of 255 represent the salient target identification results for the slump bucket and the cement, respectively. Step 3.2, the binarized image of the slump before the experiment obtained in Step 3.

1. Perform a search to find the highest y-axis coordinate value in the region with a grayscale value of 255. and lowest point Calculate the absolute difference between the two points on the y-axis coordinates, and then calculate the ratio K of the image size to the actual size based on the actual height of the slump bucket. Step 3.3: The binarized image obtained from the slump experiment in Step 3.1 is processed. Perform a search to find the point with the minimum y-axis coordinate value in the region where the grayscale value is 255. Calculate the absolute difference between the point and the minimum y-axis coordinate of the region with a gray value of 255 in P, and then calculate the actual height T of the collapse based on K. Step 3.4: Calculate the concrete classification results based on slump grading. The slump is 10-40mm, which is low plasticity concrete; 50-90mm, which is plastic concrete; 100-150mm, which is fluid concrete; and slump greater than 160mm, which is highly fluid concrete. Step 3.5, output P. T and concrete classification results.

2. The concrete slump detection method based on salient target detection according to claim 1, characterized in that, The image preprocessing step 1 is specifically implemented according to the following steps: Step 1.1: Collect concrete sample images with different slump and label them, marking the collapsed areas as salient targets to construct a labeled training set; Step 1.2: Process the images in the training dataset to generate similar but not identical samples for data augmentation. Step 1.3: Normalize the pixel values ​​of the image to a specific range.

3. The concrete slump detection method based on salient target detection according to claim 1, characterized in that, Step 2.2 is implemented in the following steps: Step 2.2.1: Construct a network with ResNet50 as the backbone; Step 2.2.2: Multi-scale feature extraction is achieved by stacking multiple residual blocks; Step 2.2.3: After the last residual block, a global average pooling layer is added to perform average pooling on the feature map of each channel to obtain a feature vector of fixed length.

4. The concrete slump detection method based on salient target detection according to claim 1, characterized in that, Step 2.3 is implemented in the following steps: Step 2.3.1: Input the basic features extracted from the five stages of the ResNet50 network. ~ As the input to the first bidirectional dense feature aggregation module, this module consists of three bidirectional dense feature aggregation layers. Step 2.3.2: The first bidirectional dense feature aggregation layer takes five signals as input and outputs: four signals as input to the next layer and the intermediate multi-scale semantic features. ; Step 2.3.3: The second bidirectional dense feature aggregation layer takes four signals as input and outputs the three signals that will serve as input to the next layer, along with the intermediate multi-scale semantic features. ; Step 2.3.4, the third bidirectional dense feature aggregation layer, takes three signals as input and outputs three intermediate multi-scale semantic features. ; Step 2.3.5: Combine the five intermediate multi-scale semantic features The input is fed into the next bidirectional dense feature aggregation module. The two bidirectional dense feature aggregation modules together form a cross-bidirectional dense feature aggregation module with the same structure. The final output is the final multi-scale semantic feature. , .

5. The concrete slump detection method based on salient target detection according to claim 4, characterized in that, Step 2.3.2 is implemented in the following steps: Step 2.3.2.1, Through upsampling and Perform element-wise linear weighting and convolution operations to obtain Similarly, obtain them sequentially from bottom to top. , as well as ; Step 2.3.2.2, downsampling and , Perform element-wise linear weighting and convolution operations to obtain Similarly, obtain them sequentially from top to bottom. , , ;in As one of the final multi-scale semantic features ,at the same time , , , As the input to the next bidirectional dense feature aggregation layer, the fusion formula is as follows: （1） （2） In the formula, conv represents the convolution operation on the features, which consists of convolution operation, batch normalization, and ReLU activation function; The weight coefficients assigned to each input during the feature fusion stage are initialized to random numbers in the range (0,1).

6. The concrete slump detection method based on salient target detection according to claim 4, characterized in that, Step 2.4, constructing the multi-resolution semantic complementarity module, is implemented according to the following steps: The final multi-scale semantic features output by the bidirectional dense feature aggregation module in steps 2.4.1 and 2.

3. , As input to the multi-resolution semantic complementarity module, these features are halved in resolution and contain semantic information from targets at different scales. Step 2.4.2 involves concatting two adjacent input features, followed by dilated convolution with a rate of 2, normalization, and ReLU operations to obtain the global semantic information that fuses the two input features; the formula is as follows: （3） In the formula, , conv represents adjacent features, and conv represents holes with batch normalization layers and ReLU activation functions; Step 2.4.3 involves adding the obtained global semantic information back into the input features through element-wise summation, thereby enhancing the input features both spatially and semantically; the formula is as follows: （4） （5） In the formula, This is element-wise addition; Step 2.4.4 involves selectively inheriting two sets of spatially and semantically reinforced features through element-wise multiplication. This achieves the goal of allowing the input features to inherit important characteristics while discarding more noise. The formula is as follows: （6） In the formula, This is element-wise multiplication.

Images