A method for inspecting the quality of ground leveling by a concrete construction robot

Abstract

This application relates to a method for detecting the surface leveling quality of concrete construction robots, comprising the following steps: Step S1, constructing a bilateral segmentation network, including constructing a detail branch and a semantic segmentation branch; Step S2, generating the detail branch by stacking WSBlocks; Step S3, capturing the semantic information of the concrete surface image by stacking NDBlocks; Step S4, adding a multi-scale wavelet feature fusion network (WFFN) to the detail branch; Step S5: inserting an illumination-adaptive feature extraction module into the shallow part of the semantic segmentation branch; Step S6: fusing deep and shallow features using a bilateral guided aggregation layer, aligning the last layer features of the two branches through upsampling and downsampling to obtain fused features, thereby improving the accuracy of concrete defect segmentation. The efficient, accurate, and intelligent concrete leveling defect detection technology researched in this invention solves the problems of concrete construction robots being unable to perform online leveling quality detection and having low detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly provides a concrete construction robot ground leveling quality detection method. BACKGROUND

[0002] In a construction project, the concrete leveling quality directly affects the safety, durability of the structure and the construction effect of the subsequent decoration project. Tradically, the detection of concrete leveling defects (such as uneven surface, cracks, hollowing, sanding, etc.) mainly relies on manual methods, such as partial sampling detection by tools such as rulers and levels. However, these methods have limitations such as low efficiency, strong subjectivity, difficulty in covering large areas, and limited ability to identify hidden defects.

[0003] In recent years, with the improvement of construction quality requirements in the construction industry and the development of intelligent technology, advanced methods based on computer vision detection have been gradually applied to concrete leveling defect detection. This method mainly includes two categories: traditional image processing and deep learning. Traditional image processing mainly includes edge detection, threshold segmentation, texture analysis, and morphological processing. Deep learning mainly includes detection algorithms and segmentation algorithms. In terms of concrete surface defect detection, deep learning-based methods can detect defects more quickly and accurately.

[0004] However, the above methods are all for defects in solidified concrete, and there are still differences with defects in freshly poured concrete surfaces. In addition, there are problems such as light, blurred defect edges, etc. in the actual detection process, which leads to the fact that existing algorithms cannot achieve high-precision real-time detection, and the detection of concrete ground leveling quality is still a challenging research topic.

[0005] Therefore, it is of great significance to study efficient, accurate and intelligent concrete leveling defect detection technology to improve construction quality and reduce rework costs. SUMMARY

[0006] In view of this, the purpose of the present application is to provide a concrete construction robot ground leveling quality detection method to solve the problems of the concrete construction robot being unable to perform online leveling quality detection and low detection accuracy.

[0007] A concrete construction robot ground leveling quality detection method, the improvement lies in comprising the following steps:

[0008] Step S1, a bilateral segmentation network is constructed, including constructing a detail branch and constructing a semantic segmentation branch; the detail branch is used to process edge information; the semantic segmentation branch is used to process semantic information;

[0009] Step S2: Generate detail branches by stacking WSBlocks; the detail branches focus on the details of the underlying concrete layer and are used to capture shallow information of the concrete surface.

[0010] Step S3: Capture semantic information of concrete ground images by stacking NDBlock; the first layer of NDBlock is an illumination-adaptive feature extraction module; the middle three layers are lightweight convolutions; the last layer is a dynamic region attention mechanism, so that the semantic segmentation branch can finally obtain a 16×16 feature map;

[0011] Step S4: Add the multi-scale wavelet feature fusion network (WFFN) to the detail branch to enhance and fuse the high-frequency features of the concrete image at each scale, thereby making the detail branch more prominent for concrete leveling defects.

[0012] Step S5: Insert the illumination adaptive feature extraction module into the shallow part of the semantic segmentation branch. Dynamically predict illumination compensation parameters through lightweight convolution to perform pixel-level correction on the input feature map. Also add a dynamic region attention mechanism to the last layer of the semantic segmentation branch and dynamically adjust the attention weight according to the local brightness of the image to reduce the response of high brightness areas, thereby reducing the impact of illumination on concrete defect detection.

[0013] Step S6: Use a bilateral guided aggregation layer to fuse deep and shallow features. Align the last layer features of the two branches with upsampling and downsampling to obtain fused features, thereby improving the accuracy of concrete defect segmentation.

[0014] Preferably, in step S1, after the input image enters the network, the edge information of the input image is obtained through the detail branch, and the edge information includes the spatial location features of the defects; at the same time, after the input image enters the network, the semantic information of the input image is obtained through the semantic segmentation branch, and the semantic information includes the shape and category features of the defects.

[0015] Preferably, WSBlock uses two ordinary convolutions and one deformable convolution to extract features from the concrete ground image. The detail branch includes three WSBlock layers to generate a 128-channel feature map.

[0016] Preferably, the deformable convolutional block is used to change the offset of each sampling point, adopting an irregular sampling form that offsets in any direction, as shown in the following formula:

[0017]

[0018] Where p0 represents any point in the image, p n R represents the offset of each point in the convolution kernel relative to the center point, where R is a real number, and ω(p) n) represents the weight at the corresponding position of the convolution kernel, x(p0+p n ) represents the element value at position on the input feature map, Δp n This indicates the offset.

[0019] Preferably, step S4 includes:

[0020] Step S4-1: Convert the input image to grayscale and perform wavelet decomposition to obtain low-frequency and high-frequency components. Then enhance the high-frequency sub-band and finally reconstruct the image.

[0021] Wavelet transform involves generating a set of wavelet bases through translation and scaling of the mother wavelet, as shown in the following formula:

[0022]

[0023] Where a is the scale parameter, b is the translation parameter, and t is the sampling period;

[0024] The discrete wavelet transform decomposes the signal x(t) as follows:

[0025]

[0026] Step S4-1-1, perform wavelet decomposition on the input signal, including: perform DWT on the input signal x to obtain the low-frequency (A) component A=x*φ and the high-frequency (D) component D=x*ψ, where φ is the scaling function and ψ is the wavelet function;

[0027] Step S4-1-2: Apply an independent convolution kernel K to each subband. A and K D :

[0028] y A =A*K A

[0029] y D =D*K D ;

[0030] Step S4-1-3, finally merge the results y using inverse wavelet transform:

[0031] y = IDWT(y A ,y D );

[0032] Step S4-2: Perform wavelet decomposition on each layer in the detail branch, and then fuse the three-layer high-frequency enhanced feature map with the final feature map to highlight the leveling defects of the concrete floor.

[0033] Preferably, step S5 includes:

[0034] Step S5-1: The illumination adaptive feature extraction module is composed of illumination adaptive convolution. The illumination adaptive convolution captures the overall brightness through global average pooling, and then reduces the dimensionality of the overall brightness to reduce the computational load, outputting the illumination intensity α and the ambient illumination offset β.

[0035] Step S5-2: In each layer of the semantic segmentation branch, an illumination-adaptive feature extraction module is embedded. The extracted features are fused with the input feature map of the previous layer and then sent to the next layer for feature extraction.

[0036] Step S5-3: The dynamic region attention mechanism first generates a single-channel attention map, then calculates the luminance mask, and dynamically adjusts the attention based on the result to enhance the defect features.

[0037] Preferably, the illumination-adaptive feature extraction module includes a main convolutional path and an illumination prediction branch; the illumination linear model used in the illumination-adaptive feature extraction module is:

[0038] I(x,y)=α*R(x,y)+β

[0039] Where I(x,y) is the pixel value observed after being affected by illumination; R(x,y) is the true reflectivity; α is the illumination intensity (α>0); and β is the ambient illumination offset.

[0040] The semantic segmentation branch extracts global statistics through the illumination prediction branch, as shown in the following formula:

[0041]

[0042] Where H and W are the image length and width; I c (x, y) is I(x, y) at point c;

[0043] The predicted parameters α and β are obtained through a fully connected layer, and pixel-wise correction is performed as shown in the following equation:

[0044] I corrected (x,y)=σ(α)·I(x,y)+β.

[0045] Preferably, the dynamic region attention mechanism consists of a luminance mask branch and a spatial attention branch; the output is the input feature map, luminance mask, and attention. Figure Three The dot product is used to obtain the luminance mask; the luminance mask is obtained by comparing the grayscale value with the luminance threshold, and the formula is:

[0046]

[0047] Among them, I gray (x,y) represents the grayscale value; T=σ(t) is constrained to (0.5,0.9) by the Sigmoid function;

[0048] The spatial attention branch obtains the basic attention through a 1×1 convolution, then undergoes Gaussian smoothing, and finally sums with M. brightness The dot product of (x,y) yields the dynamic region attention output.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention provides a method for inspecting the ground leveling quality of concrete construction robots. This method constructs an integrated model for concrete leveling defect detection and localization, effectively enabling the detection of abnormal concrete leveling defects, the determination of defect types, and defect localization. This model allows for real-time detection on the concrete construction robot, improving the robot's autonomy. Attached Figure Description

[0051] Figure 1 This is a diagram illustrating the overall algorithm framework provided in the embodiments of the present invention.

[0052] Figure 2 A schematic diagram of deformable convolution provided in the embodiments of the present invention;

[0053] Figure 3 This is a schematic diagram of wavelet convolution W-Block provided in the embodiments of the present invention;

[0054] Figure 4 This is a structural diagram of the illumination adaptive feature extraction module provided in an embodiment of the present invention;

[0055] Figure 5 This is a structural diagram of the dynamic region attention mechanism provided in the embodiments of the present invention. Detailed Implementation

[0056] To better understand this invention, the following description, in conjunction with the accompanying drawings and examples, will further illustrate the invention.

[0057] The technical solution of the present invention will be further described in detail below.

[0058] A bilateral guided aggregation layer is a module used in visual tasks (such as semantic segmentation and image enhancement) to fuse and enhance feature representation capabilities. This application proposes a method for detecting the ground leveling quality of a concrete construction robot based on a bilateral guided aggregation layer. The improvement lies in the following steps:

[0059] Step S1: Construct a bilateral segmentation network, including constructing a detail branch and a semantic segmentation branch; the detail branch is used to process edge information; the semantic segmentation branch is used to process semantic information.

[0060] Specifically, in step S1, after the input image enters the network, its edge information, including the spatial location features of defects, is obtained through the detail branch; simultaneously, its semantic information, including the shape and category features of defects, is obtained through the semantic segmentation branch. For example... Figure 1 As shown, after the input image enters the network, it is fed into the Detail branch and the Segmentation branch to extract features respectively. The Detail branch has fewer convolutional layers, and the features obtained are more focused on the spatial location of defects, while the Segmentation branch has more convolutional layers, and the features obtained are more focused on the shape and category of defects. Designing a dual-branch architecture and constructing a bilateral segmentation network, extracting image features through two branches with different structures, can accelerate the network feature extraction speed and reduce complexity.

[0061] Step S2: Detail branches are generated by stacking WSBlocks; the detail branches focus on the details of the underlying concrete layer and are used to capture shallow information of the concrete surface.

[0062] WSBlock employs a wide-channel, shallow-layer structure to obtain feature information with a large spatial dimension, thereby quickly acquiring edge information and spatial details of concrete. The detail branches generated by stacking WSBlocks, with their wide-channel, shallow-layer characteristics, focus only on the underlying details, thus capturing shallow information of the concrete surface. These underlying details include the shape and texture of concrete defects; the shallow information includes the boundary contours and location information of the defects.

[0063] WSBlock employs two ordinary convolutions and one deformable convolution to extract features from concrete surface images. The detail branch consists of three WSBlock layers, generating a 128-channel feature map. Specifically, WSBlock comprises three convolutional blocks: two ordinary convolutional blocks and one deformable convolutional block. The image input dimension is [3, 512, 512]. After the first convolutional layer, a feature map with dimensions [64, 256, 256] is obtained; after the second convolutional layer, a feature map with dimensions [64, 128, 128] is obtained; and after the final convolutional layer, a feature map with dimensions [128, 64, 64] is obtained, ultimately resulting in a 64×64 feature map of concrete defects. This approach more easily preserves the spatial features of defects and focuses on their edge positions.

[0064] The deformable convolutional block is used to change the offset of each sampling point, adopting an irregular sampling form that offsets in any direction, as shown in the following formula:

[0065]

[0066] Where p0 represents any point in the image, p n ω(p) represents the offset of each point in the convolution kernel relative to the center point. n ) represents the weights at the corresponding positions of the convolution kernel, R is a real number, and x(p0+p n ) represents the element value at position on the input feature map, Δp n This indicates the offset.

[0067] Specifically, in ordinary convolution, the kernels are all 3×3, the first convolution has a stride of 1, and the second convolution has a stride of 2. Both convolutional blocks are composed of Conv2d+BN+ReLU. Variable convolution differs from ordinary convolution in that it better considers the shape variations of the detected target, enabling the network to have greater tolerance for concrete defect shapes. For example... Figure 2 As shown, deformable convolution changes the offset of each sampling point on the basis of ordinary convolution, so that it is no longer a conventional matrix sampling, but an irregular sampling form that can be offset in any direction.

[0068] Step S3 captures high-level semantic information of concrete ground images by stacking NDBlocks, which can also be understood as stacking NDBlocks to obtain deep features of concrete; the first layer of the NDBlock is an illumination-adaptive feature extraction module; the middle three layers are lightweight convolutions; the last layer is a dynamic region attention mechanism, so that the semantic segmentation branch can finally obtain a 16×16 feature map.

[0069] Specifically, such as Figure 1 As shown, NDBlock employs a narrow-channel, deep structure to capture high-level semantic information of concrete surface images. The first layer is an illumination-adaptive feature extraction module; the middle three layers are lightweight convolutional structures to reduce the number of parameters and improve feature acquisition speed; the last layer is a dynamic region attention mechanism, which ultimately yields a 16×16 feature map.

[0070] Step S4: Add the multi-scale wavelet feature fusion network (WFFN) to the detail branch to enhance and fuse the high-frequency features of the concrete image at each scale, thereby making the detail branch more prominent for concrete leveling defects.

[0071] Specifically, such as Figure 3 As shown, wavelet transform is constructed into a wavelet convolutional layer WTConv, and then a multi-scale wavelet feature fusion network WFFN is added to the detail branch to enhance and fuse the high-frequency features of the concrete image at each scale, thereby highlighting concrete leveling defects more effectively through the detail branch.

[0072] Step S4-1: Convert the input image to grayscale and perform wavelet decomposition to obtain low-frequency and high-frequency components. Then, enhance the high-frequency sub-band and finally reconstruct the image.

[0073] Wavelet transform involves generating a set of wavelet bases through translation and scaling of the mother wavelet, as shown in the following formula:

[0074]

[0075] Where a is the scale parameter, b is the translation parameter, and t is the sampling period;

[0076] The discrete wavelet transform decomposes the signal x(t) as follows:

[0077]

[0078] The traditional convolution kernel is replaced with a wavelet basis function, or wavelet transform is introduced during the convolution process. Step S4-1-1, wavelet decomposition is performed on the input signal, including: performing DWT on the input signal x to obtain the low-frequency (A) component A=x*φ and the high-frequency (D) component D=x*ψ, where φ is the scaling function and ψ is the wavelet function.

[0079] Step S4-1-2: Apply an independent convolution kernel K to each subband. A and K D :

[0080] y A =A*K A

[0081] y D =D*K D ;

[0082] Step S4-1-3, finally merge the results y using inverse wavelet transform:

[0083] y = IDWT(y A ,y D ).

[0084] Step S4-2: Perform wavelet decomposition on each layer in the detail branch, and then fuse the three-layer high-frequency enhanced feature map with the final feature map to highlight the leveling defects of the concrete floor.

[0085] Specifically, after constructing the wavelet convolution, a WTConv is embedded in the output of each layer in the detail branch, and the output is fused with the original features before being fed into the next layer. Here, the original features refer to the feature maps output by each WSBlock layer. For example... Figure 1 As shown, by enhancing the frequency domain perception capability of convolution operations through the mathematical framework of wavelet transform, the location of defects in concrete leveling can be better distinguished, which alleviates the problem of low contrast between the foreground and background during detection to some extent.

[0086] Step S5: Insert the illumination adaptive feature extraction module into the shallow part of the semantic segmentation branch. Dynamically predict illumination compensation parameters through lightweight convolution to perform pixel-level correction on the input features. Also add a dynamic region attention mechanism to the last layer of the semantic segmentation branch. Dynamically adjust the attention weights according to the local brightness of the image to reduce the response of high-brightness areas, thereby reducing the impact of illumination on concrete defect detection.

[0087] Step S5-1: The illumination adaptive feature extraction module consists of illumination adaptive convolution. The illumination adaptive convolution captures the overall brightness through global average pooling, then reduces its dimensionality to reduce computation, and outputs the illumination intensity α and the ambient illumination offset β. Sigmoid is used on α to prevent overexposure.

[0088] Specifically, such as Figure 4 As shown, the illumination adaptive feature extraction module consists of a main convolutional path and an illumination prediction branch. The illumination linear model is as follows:

[0089] I(x,y)=α*R(x,y)+β;

[0090] Where I(x,y) is the observed pixel value (after being affected by illumination); R(x,y) is the true reflectivity (an inherent property of the object, such as defect texture); α is the illumination intensity (brightness gain, >0); and β is the ambient illumination offset (contrast offset).

[0091] The network extracts global statistics through the illumination prediction branch, as shown in the following formula:

[0092]

[0093] Where H and W are the image length and width; I c (x, y) represents I(x, y) at point c, which is the pixel value observed at point c after being affected by illumination.

[0094] The predicted parameters α and β are obtained through a fully connected layer, and pixel-wise correction is performed as shown in the following equation:

[0095] I corrected (x,y)=σ(α)·I(x,y)+β;

[0096] Step S5-2: In each layer of the semantic segmentation branch, an illumination-adaptive feature extraction module is embedded. The extracted features are fused with the input feature map of the previous layer and then sent to the next layer for further feature extraction.

[0097] Step S5-3: The dynamic region attention mechanism first generates a single-channel attention map, then calculates the luminance mask, and dynamically adjusts the attention based on the result to enhance the defect features.

[0098] Specifically, such asFigure 5 As shown, the dynamic region attention mechanism consists of a luminance mask branch and a spatial attention branch. The output includes the input feature map, the luminance mask, and the attention function. Figure Three The luminance mask is obtained by multiplying the grayscale values ​​by the luminance threshold. The formula is:

[0099]

[0100] Among them, I gray (x,y) represents the grayscale value; T = σ(t) is constrained to (0.5, 0.9) by the Sigmoid function.

[0101] The spatial attention branch obtains the basic attention through a 1×1 convolution, then undergoes Gaussian smoothing, and finally sums with M. brightness The dot product of (x,y) yields the dynamic region attention output.

[0102] Step S6: Use a bilateral guided aggregation layer to fuse deep and shallow features. Align the last layer features of the two branches with upsampling and downsampling to obtain fused features, thereby improving the accuracy of concrete defect segmentation.

[0103] Experiments have shown that the ground leveling quality detection method for concrete construction robots proposed in this invention has an accuracy rate of 96%, a positioning accuracy of 89%, and a detection speed of 0.43s, enabling real-time detection on concrete construction robots.

[0104] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure One One or more processes and / or boxes Figure One A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure One One or more processes and / or boxes Figure One The function specified in one or more boxes.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure One One or more processes and / or boxes Figure One The steps of the function specified in one or more boxes.

[0108] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for detecting the quality of ground leveling by a concrete construction robot, characterized by, The method comprises the following steps: Step S1, constructing a bilateral segmentation network, including constructing a detail branch and constructing a semantic segmentation branch; the detail branch is used for processing edge information; the semantic segmentation branch is used for processing semantic information; Step S2, generating the detail branch by stacking WSBlock; the detail branch focuses on the details of the concrete bottom layer and is used for capturing the shallow information of the concrete surface; Step S3, capturing the semantic information of the concrete ground image by stacking NDBlock; the first layer of the NDBlock is a light adaptive feature extraction module; The middle three layers are light convolution; and the last layer is a dynamic region attention mechanism, so that the semantic segmentation branch can finally obtain a 16x16 feature map; Step S4, adding a multi-scale wavelet feature fusion network WFFN to the detail branch, enhancing and fusing the high-frequency features of the concrete image at each scale, so that the detail branch can highlight the concrete leveling defects; Step S5: inserting the light adaptive feature extraction module into the shallow part of the semantic segmentation branch, dynamically predicting the light compensation parameters through light adaptive convolution, performing pixel-level correction on the input feature map, adding a dynamic region attention mechanism to the last layer of the semantic segmentation branch, and dynamically adjusting the attention weight according to the local brightness of the image to reduce the response of the high-brightness area, thereby reducing the influence of light on the detection of concrete defects; Step S6: adopting a bilateral guided aggregation layer to fuse deep and shallow features, aligning the last layer features of the two branches through upsampling and downsampling to obtain fused features, and improving the segmentation accuracy of concrete defects; The step S4 comprises: Step S4-1: converting the input image into a grayscale image and performing wavelet decomposition to obtain low-frequency components and high-frequency components, then enhancing the high-frequency subband, and finally reconstructing the image; Wavelet transform comprises generating a group of wavelet bases through translation and stretching of a mother wavelet, and the formula is as follows: ； wherein, is a scale parameter, is a translation parameter; t is a sampling period; Discrete wavelet transform for signal decomposition formula is: ； Step S4-1-1, wavelet decomposition of the input signal, comprising: performing DWT on the input signal to obtain a low-frequency component , and a high-frequency component , wherein, is a scale function, is a wavelet function; Step S4-1-2, applying an independent convolution kernel to each sub-band and : ； Step S4-1-3, finally the results are merged by inverse wavelet transform : ； Step S4-2: performing wavelet decomposition on each layer in the detail branch, then fusing the three layers of high-frequency enhanced feature maps with the final feature map, so as to highlight the concrete ground leveling defects.

2. The concrete construction robot ground finishing quality detection method according to claim 1, characterized by, In step S1, the edge information of the input image is obtained through the detail branch after the input image enters the network, and the semantic information of the input image is obtained through the semantic segmentation branch after the input image enters the network.

3. The concrete construction robot ground finishing quality detection method according to claim 1, characterized by, The WSBlock adopts 2 ordinary convolutions and 1 deformable convolution to jointly extract features of the concrete ground image, and the detail branch comprises 3 WSBlock layers to generate a 128-channel feature map.

4. The concrete construction robot ground finishing quality detection method according to claim 3, characterized by, The deformable convolution is used to change the offset of each sampling point, and adopts an irregular sampling form that offsets in any direction, as shown in the following formula: ； wherein, represents an arbitrary point in the image, represents an offset of each point in the convolution kernel relative to the center point, R belongs to real numbers, represents the weight of the corresponding position of the convolution kernel, represents the element value at the position on the input feature map, represents the offset.

5. The concrete construction robot ground finishing quality detection method according to claim 1, characterized by, The step S5 comprises: Step S5-1, the light adaptive feature extraction module is composed of a light adaptive convolution, which captures the overall brightness through global average pooling, and then reduces the dimension of the overall brightness to reduce the calculation amount, and outputs the light intensity a and the ambient light offset β; Step S5-2: An illumination adaptive feature extraction module is embedded in each layer in the semantic segmentation branch, and the extracted features are fused with the input feature maps of the previous layer and then sent to the next layer for feature extraction; Step S5-3: The dynamic region attention mechanism first generates a single-channel attention map, then calculates a brightness mask, dynamically adjusts the attention according to the result, and thus enhances the defect features.

6. The concrete construction robot ground finishing quality detection method according to claim 1, characterized by, The illumination adaptive feature extraction module includes a main convolution path and an illumination prediction branch; the illumination linear model used by the illumination adaptive feature extraction module is as follows: ； wherein, is the observed pixel value after being affected by the illumination; is the true reflectance; is the intensity of the illumination > 0); is the ambient light illumination offset; The semantic segmentation branch extracts global statistics through the illumination prediction branch, as shown in the following formula: ； wherein, is the image length and width; I c (x, y) is I(x, y) at c; get the predicted parameters through the fully connected layer pixel-wise correction is performed as follows: 。 7. The concrete construction robot ground finishing quality detection method according to claim 1, characterized by, The dynamic region attention mechanism is composed of a brightness mask branch and a spatial attention branch; the output result is obtained by point multiplication of the input feature maps, the brightness mask and the attention map; The brightness mask is obtained by comparing the gray value with the brightness threshold, and the formula is as follows: ； wherein, is a gray value; constrained to (0.5, 0.9) by Sigmoid. The spatial attention branch gets the base attention through 1x1 convolution, then goes through a Gaussian smoothing, and finally gets the dynamic region attention output by point multiplication with the base attention. the base attention.

Images