Method, device and equipment for identifying segregation of asphalt mixture

Abstract

The application provides a kind of asphalt mixture segregation identification method, device and equipment, it is related to segregation identification technical field.The method comprises: the image preprocessing is carried out to the collected asphalt mixture image to be identified, and the preprocessed image to be identified is obtained;Based on the preset low-illumination image enhancement algorithm, the preprocessed image to be identified is carried out image enhancement processing, and the standard image of asphalt mixture is obtained;The standard image is input into the convolutional neural network model trained in advance, and the segregation identification result of asphalt mixture is obtained;Wherein, the convolutional neural network model trained in advance includes general feature extraction module, the feature extraction module of convolution unit including different sizes of convolution kernel, the pooling module including multistage pooling structure, fully connected module and output module.The application can accurately capture the feature difference of different segregation degrees from the low-illumination asphalt mixture image, and improve the accuracy of segregation identification.

Description

Technical Field

[0001] This invention relates to the field of segregation identification technology, and in particular to a method, apparatus and equipment for identifying segregation in asphalt mixtures. Background Technology

[0002] Asphalt mixtures are a core material in road construction, and the uniformity of their mixing and paving directly determines the load-bearing capacity and service life of the pavement structure. Asphalt mixture segregation is a common quality defect during construction. Segregated asphalt mixtures have unbalanced aggregate gradation and uneven asphalt content distribution, easily leading to early cracks, potholes, and water damage, significantly reducing road service life and increasing maintenance costs. The inventors have found that most existing asphalt mixture segregation identification methods are designed for well-lit conditions. In actual low-light environments such as nighttime or low-light conditions, their segregation identification effect is poor and their detection efficiency is low. Summary of the Invention

[0003] This invention provides a method, apparatus, and equipment for identifying asphalt mixture segregation, to solve the problem that existing technologies have poor segregation identification effects and low detection efficiency in low-light environments such as nighttime or low-light conditions.

[0004] In a first aspect, embodiments of the present invention provide a method for identifying asphalt mixture segregation, comprising: The collected images of asphalt mixtures to be identified are preprocessed to obtain the preprocessed images to be identified. Based on a preset low-light image enhancement algorithm, the preprocessed image to be identified is enhanced to obtain a standard image of asphalt mixture. A standard image is input into a pre-trained convolutional neural network model to obtain the segregation identification result of asphalt mixture. The pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module. The feature extraction module includes convolutional units with convolutional kernels of different sizes, and the pooling module includes a multi-level pooling structure.

[0005] Secondly, embodiments of the present invention provide an asphalt mixture segregation identification device, comprising: The preprocessing module is used to preprocess the acquired asphalt mixture images to be identified, and obtain the preprocessed images to be identified. The enhancement module is used to perform image enhancement processing on the preprocessed image to be identified based on a preset low-light image enhancement algorithm to obtain a standard image of asphalt mixture; The recognition module is used to input standard images into a pre-trained convolutional neural network model to obtain the segregation recognition results of asphalt mixtures. The pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module. The feature extraction module includes convolutional units with convolutional kernels of different sizes, and the pooling module includes a multi-level pooling structure.

[0006] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0007] In this embodiment of the invention, the acquired image of asphalt mixture to be identified is preprocessed to obtain a preprocessed image. Then, according to a preset low-light image enhancement algorithm, the preprocessed image is enhanced to obtain a standard image of asphalt mixture. This optimizes the image quality of the asphalt mixture image and enhances the recognizability of segregation features in the asphalt mixture image. In addition, the pre-trained convolutional neural network model includes a feature extraction module with convolutional kernels of different sizes, a pooling module with multi-level pooling structure, a fully connected module, and an output module. Using convolutional kernels of different sizes for feature extraction can improve the stability of feature distribution and nonlinear expression ability, thereby enhancing the model's representation effect on complex target features. Using multi-level pooling structure for pooling can make the feature map size decrease step by step and in an orderly manner with the convolutional unit, strengthening key local features while retaining overall feature information, thereby improving feature selection ability and model computation efficiency. Therefore, using the convolutional neural network model for segregation recognition can accurately capture the feature differences of different degrees of segregation from low-light asphalt mixture images, improving the accuracy of segregation recognition. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the implementation of the asphalt mixture segregation identification method provided in this embodiment of the invention. Figure 2 This is a model architecture diagram of a pre-trained convolutional neural network model for the asphalt mixture segregation identification method provided in this embodiment of the invention; Figure 3 This is a flowchart illustrating the implementation of step S130 of the asphalt mixture segregation identification method provided in this embodiment of the invention. Figure 4 This is a schematic diagram of the asphalt mixture segregation identification device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0009] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0010] See Figure 1 The flowchart illustrating the implementation of the asphalt mixture segregation identification method provided in this embodiment of the invention is described in detail below: Step S110: Perform image preprocessing on the acquired image of the asphalt mixture to be identified to obtain the preprocessed image to be identified.

[0011] In some embodiments, image preprocessing includes image size normalization and grayscale conversion. Image size normalization refers to uniformly adjusting the acquired asphalt mixture images to a fixed size. This size can be set according to the input requirements of the subsequent convolutional neural network model. Image size normalization ensures that all input images have consistent dimensions, avoiding the impact of size differences on model training performance.

[0012] It should be noted that grayscale conversion refers to using a weighted average method to calculate the weighted average of the RGB pixel values ​​of an image to be recognized, adjusted to a fixed size, to obtain a single-channel grayscale image. By performing grayscale conversion on the image to be recognized, adjusted to a fixed size, the main texture and brightness information of the mixture surface can be effectively preserved while reducing data dimensionality. The formula for the weighted average calculation is as follows:

[0013] in, The weighted pixel grayscale values. , , The pixel values ​​in the red, green, and blue channels of the image to be recognized, which has been resized to a fixed size. , , These are the weighting coefficients for the red, green, and blue channels.

[0014] Step S120: Based on the preset low-light image enhancement algorithm, perform image enhancement processing on the preprocessed image to be identified to obtain a standard image of asphalt mixture.

[0015] In some embodiments, the preset low-light image enhancement algorithm can be the Zero-DCE (Zero-Reference Deep Curve Estimation) algorithm. The Zero-DCE algorithm can enhance the brightness of the preprocessed image to be recognized by learning a pixel-level nonlinear mapping curve, and achieve adaptive enhancement of the preprocessed low-light image to be recognized without the need for a reference high-quality image.

[0016] In one possible implementation, step S120 is specifically processed as follows: an adaptive image enhancement formula is determined according to a preset low-light image enhancement algorithm; based on the adaptive image enhancement formula, the pixel grayscale value at each position of the preprocessed image to be identified is enhanced to obtain a standard image of the asphalt mixture.

[0017] In some embodiments, when the preset low-light image enhancement algorithm is the Zero-DCE algorithm, the adaptive image enhancement formula is:

[0018] in, The standard image after image enhancement in coordinates The pixel grayscale value at that location, To preprocess the image to be recognized in coordinates The pixel grayscale value at that location, This is the set of pixel-level enhancement curve parameters predicted by a convolutional neural network. The enhancement function is composed of multiple pointwise nonlinear mappings. The expression is:

[0019] in, To preprocess the image to be recognized in coordinates The pixel grayscale value at that location, For the first The curve adjustment parameter of the augmentation mapping is obtained by end-to-end learning of the convolutional neural network.

[0020] Step S130: Input the standard image into the pre-trained convolutional neural network model to obtain the segregation identification result of the asphalt mixture; wherein, the pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module; wherein, the feature extraction module includes convolutional units with convolutional kernels of different sizes; the pooling module includes a multi-level pooling structure.

[0021] In some embodiments, the pre-trained convolutional neural network model needs to be constructed using transfer learning. The pre-trained convolutional neural network model uses a ResNet50 (Residual Network 50 layers) model as the base model, freezes the parameters of the first 10 layers of the ResNet50 model, and improves the subsequent model architecture. This can avoid destroying the general features of the ResNet50 model, make full use of the general image features already learned by the model, reduce the training difficulty of the model of the present invention, and improve the recognition performance.

[0022] It should be noted that the model architecture of a pre-trained convolutional neural network model is as follows: Figure 2 As shown, the pre-trained convolutional neural network model includes modules such as a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module. The feature extraction module includes multi-level cascaded convolutional units. These units use convolutional kernels of different sizes and progressively increase the number of channels, enabling multi-scale extraction of fine-grained texture features and local contour features. It also combines batch normalization and ReLU (Rectified Linear Unit) activation to improve the stability of feature distribution and non-linear expressive power, thereby enhancing the model's representation of complex target features. The pooling module includes a multi-level pooling structure, employing a hybrid pooling strategy combining max pooling and average pooling. This allows the feature map size to decrease progressively and orderly with each convolutional unit, strengthening key local features while preserving overall feature information, thus improving feature selection capability and model computational efficiency.

[0023] See Figure 3 The specific processing method of step S130 above includes steps S1301-S1305, the details of which are as follows: Step S1301: Input the standard image into the general feature extraction module to obtain the general visual feature map of the asphalt mixture.

[0024] In some embodiments, the general feature extraction module is the model architecture corresponding to the first 10 layers of the ResNet50 model. The general feature extraction module can be used to extract the general visual features of the asphalt mixture and obtain the general visual feature map of the asphalt mixture.

[0025] Step S1302: Input the general visual feature map into the feature extraction module to obtain the multi-scale feature map of the asphalt mixture.

[0026] In some embodiments, the feature extraction module includes a first convolutional unit, a second convolutional unit, and a third convolutional unit, with different kernel sizes for the three units. Each of the first, second, and third convolutional units in the feature extraction module is followed by a batch normalization layer and a ReLU activation function layer. The batch normalization layer standardizes the convolutional output feature map, making the feature distribution more uniform and eliminating the interference of differences in feature distribution between different batches on model training. The ReLU activation function layer introduces non-linear factors, enhancing the model's ability to fit complex features.

[0027] In one possible implementation, step S1302 is specifically processed as follows: inputting the general visual feature map into the first convolutional unit to obtain the fine-grained texture feature map of the asphalt mixture; inputting the fine-grained texture feature map into the second convolutional unit to obtain the segregation region contour feature map of the asphalt mixture; and inputting the segregation region contour feature map into the third convolutional unit to obtain the multi-scale feature map of the asphalt mixture.

[0028] In some embodiments, the first convolutional unit uses 16 3×3 kernels with a stride of 1 and same fill mode. This setting allows for the extraction of fine-grained texture features from the asphalt mixture surface without altering the size of the general visual feature map, resulting in a fine-grained texture feature map. The second convolutional unit uses 32 5×5 kernels with a stride of 1 and same fill mode. This is used to extract broader local features from the fine-grained texture feature map, capturing the overall contour of the segregated region and obtaining a segregated region contour feature map. The third convolutional unit uses 64 3×3 kernels with a stride of 2 and valid fill mode. This reduces the feature map dimensionality while extracting deeper features from the segregated region contour feature map, resulting in a multi-scale feature map.

[0029] Step S1303: Input the multi-scale feature map into the pooling module to obtain the global average pooling feature vector of the asphalt mixture.

[0030] In some embodiments, the multi-level pooling structure includes a first pooling layer, a second pooling layer, a third pooling layer, and a global average pooling layer. For example... Figure 2 As shown, the first pooling layer is connected to the first convolutional unit and is used to pool the fine-grained texture feature map output by the first convolutional unit; the second pooling layer is connected to the second convolutional unit and is used to pool the segregation region contour feature map output by the second convolutional unit; the third pooling layer is connected to the third convolutional unit and is used to pool the multi-scale feature map; the global average pooling layer is used to receive the outputs of the first pooling layer, the second pooling layer, the third pooling layer and the feature extraction module, and to perform global average pooling on these outputs to obtain the global average pooled feature vector of the asphalt mixture.

[0031] In one possible implementation, step S1303 is specifically processed as follows: the fine-grained texture feature map is input into the first pooling layer to obtain the first dimensionality-reduced feature map of the asphalt mixture; the segregation region contour feature map is input into the second pooling layer to obtain the second dimensionality-reduced feature map of the asphalt mixture; the multi-scale feature map is input into the third pooling layer to obtain the average feature map of the asphalt mixture; the multi-scale feature map, the first dimensionality-reduced feature map, the second dimensionality-reduced feature map, and the average feature map are input into the global average pooling layer to obtain the global average pooling feature vector of the asphalt mixture.

[0032] In some embodiments, both the first and second pooling layers use a 2×2 max-pooling kernel with a pooling step size of 2 and a valid fill method. This enhances key features and further reduces the dimensionality of the feature map by preserving the maximum eigenvalues ​​of local regions. Pooling is performed on fine-grained texture feature maps and segregated region contour feature maps to obtain the first and second dimensionality-reduced feature maps. The third pooling layer uses a 3×3 average pooling kernel with a pooling step size of 2 and a valid fill method. By calculating the average eigenvalues ​​of local regions, the feature map is smoothed while preserving global feature information. Pooling multi-scale feature maps using the third pooling layer yields the average feature map of the asphalt mixture. The global average pooling layer is used to perform global average pooling on the first dimensionality-reduced feature map, the second dimensionality-reduced feature map, the average feature map, and the multi-scale feature map to obtain the global average pooled feature vector of the asphalt mixture.

[0033] Step S1304: Input the global average pooling feature vector into the fully connected module to obtain the dimensionality-reduced feature vector of the asphalt mixture.

[0034] In some embodiments, the fully connected module includes a first hidden layer and a second hidden layer. The number of neurons in the first hidden layer is greater than the number of neurons in the second hidden layer. The number of neurons in the first hidden layer is 1024 and the number of neurons in the second hidden layer is 256. By adopting a multi-layer hidden structure, the depth representation and dimensionality compression of the global average pooling feature vector can be achieved through layer-by-layer nonlinear mapping.

[0035] In one possible implementation, step S1304 is specifically processed as follows: the global average pooling feature vector is input into the first hidden layer, and the global average pooling feature vector is depth-mapped based on the neurons in the first hidden layer to obtain the mapped feature vector of the asphalt mixture; the mapped feature vector is input into the second hidden layer, and the feature dimension of the mapped feature vector is compressed based on the neurons in the second hidden layer to obtain the dimensionality-reduced feature vector of the asphalt mixture.

[0036] In some embodiments, the first hidden layer is used to perform depth mapping on the global average pooling feature vector to obtain the mapped feature vector of the asphalt mixture, and the second hidden layer is used to compress the feature dimension of the mapped feature vector and extract key information from the mapped feature vector to obtain the dimensionality-reduced feature vector of the asphalt mixture.

[0037] It should be noted that each hidden layer is followed by a Dropout layer with a dropout probability of 0.5. This reduces the dependencies between neurons by randomly dropping some neuron connections, thereby preventing overfitting during training.

[0038] Step S1305: Input the dimensionality-reduced feature vector into the output module, and use the softmax function in the output module to predict the segregation probability of the dimensionality-reduced feature vector to obtain the segregation identification result of the asphalt mixture.

[0039] In some embodiments, the number of neurons in the output module needs to be consistent with the total number of segregation categories, and the Softmax function (Soft Maximum) is used as the activation function of the output module to output the probability values ​​of each category, thereby enhancing the discrimination ability and generalization performance of the pre-trained convolutional neural network model for specific segregation features, predicting the probability of different segregation categories on the dimensionality-reduced feature vector, and obtaining the segregation recognition result.

[0040] It should be noted that the segregation identification results include segregation level and segregation region location information. Segregation levels can be divided into four categories: no segregation, slight segregation, moderate segregation, and severe segregation. Each level corresponds to a clear feature judgment criterion, which can be determined based on statistical patterns such as grayscale distribution and texture features of images of each category in the sample set. Segregation region location information is determined through a heatmap corresponding to the feature map output by the convolutional neural network model. Regions with probability values ​​higher than a preset threshold in the heatmap are marked as segregated regions. The coordinates of the upper left and lower right corners of the segregated regions can be recorded using an image coordinate system. The output segregation identification results are stored in structured data format for easy subsequent data statistics and analysis. Simultaneously, a visual image containing segregated region markings is generated, allowing staff to intuitively view the segregation situation.

[0041] In one possible implementation, the training process of the pre-trained convolutional neural network model includes: performing image preprocessing and image enhancement processing on multiple normal asphalt mixture images and multiple asphalt mixture images with different degrees of segregation, respectively, to obtain a standard image corresponding to each asphalt mixture image; using each asphalt mixture image and its corresponding standard image as a training sample, and dividing all training samples into a training set, a validation set, and a test set according to a preset ratio; dividing all shuffled training samples from the training set into multiple training batches according to a preset number; iteratively updating the model parameters of the convolutional neural network model based on the training samples and loss function in each training batch; and after each iteration, The training samples from the validation set are input into the iteratively updated convolutional neural network model to obtain the recognition result of each training sample. Based on the recognition result of each training sample and the model convergence criterion, it is determined whether the iteratively updated convolutional neural network model has converged. If the iteratively updated convolutional neural network model does not converge, the learning rate in each subsequent iteration of training is adjusted based on the learning rate decay strategy, and the model parameters of the convolutional neural network model are iteratively updated based on the adjusted learning rate, the training samples in each training batch, and the loss function. If the iteratively updated convolutional neural network model converges, the iteratively updated convolutional neural network model is tested based on the training samples in the test set to obtain the pre-trained convolutional neural network model.

[0042] In some embodiments, the asphalt mixture images used to train the pre-trained convolutional neural network model need to be acquired under low-light conditions. When acquiring the images, an industrial camera with a resolution of 20 megapixels and a frame rate of 60fps can be used as the image acquisition device, equipped with a fixed-focus lens to ensure image consistency, and a ring-shaped light source with adjustable intensity as the reference light source. During acquisition, the industrial camera needs to be fixed on a bracket, and the height of the bracket should be adjusted to maintain a perpendicular distance between the camera and the asphalt mixture surface. The distance can be flexibly selected according to the actual paving area of ​​the asphalt mixture in the construction scenario, ensuring that the acquired image completely covers the target area and has clear details.

[0043] It should be noted that, to simulate the low-light environment during image acquisition, the light intensity of the acquisition area needs to be gradually reduced under controllable conditions. Specifically, this can be achieved by reducing the brightness of the ring lighting device, increasing the distance between the camera and the lighting device, or adding light-blocking material to the lens. This allows the illuminance of the target area to gradually decrease from the normal operating illuminance to the set low-light range, while maintaining uniform illumination at each illuminance level. This avoids localized highlights or excessive shadows that could distort image details, thereby obtaining representative low-light sample images.

[0044] In some embodiments, when adjusting the camera parameters of an industrial camera, the main tuning principle is "illuminance level—camera parameters—image quality." At each preset simulated low-light level, key parameters such as camera exposure time, gain (ISO), aperture (which can remain constant under fixed lens conditions), and white balance are adjusted, and the changes in image brightness and noise corresponding to different parameter combinations are recorded. Simultaneously with image acquisition, an image sharpness evaluation function is used to detect blur in the real-time acquired images, calculating the sum of image edge gradient values. When the sum of gradient values ​​is lower than a preset threshold, the image is considered blurry, and a re-acquisition command is triggered. Acquisition is repeated until images meeting the sharpness and recognizability requirements are obtained at each simulated low-light level. The corresponding illuminance values, exposure time, gain, white balance, and supplementary light intensity are simultaneously saved, ultimately forming a standard dataset and recommended parameter range for determining low-light shooting parameters.

[0045] It should be noted that during model training, the image acquisition methods described above are required to collect images of normal asphalt mixtures under low-light conditions and images of asphalt mixtures with varying degrees of segregation. Each image type should have at least 500 samples to ensure sufficient representativeness. Image preprocessing and low-light enhancement processing are performed on each asphalt mixture image to obtain a corresponding standard image. Each asphalt mixture image and its corresponding standard image are used as a training sample. All training samples are divided into training, validation, and test sets according to a preset ratio. The training set is used for model parameter learning, the validation set is used for real-time monitoring of model training performance, and the test set is used for final performance evaluation after model training. The preset ratio can be 7:2:1 or 6:2:2.

[0046] It should be noted that, in the context of Figure 2 Before training the neural network model shown, it needs to be initialized. The weights of the convolutional kernels in the feature extraction layer are initialized using a He normal distribution. This initialization method can adapt to the ReLU activation function and avoid gradient vanishing or exploding problems in the early stage of model training. The weights of the fully connected modules are initialized using a Xavier uniform distribution to ensure that the variance of the input and output data is consistent and promotes the model to converge quickly. In addition, the bias parameters are all initialized to 0, which simplifies the initialization process while ensuring the stability of model training.

[0047] In some embodiments, the loss function includes a loss function corresponding to the training batch and a loss function corresponding to each training sample. When iteratively updating the convolutional neural network model, in each iteration, the order of all training samples in the training set needs to be shuffled, and all training samples need to be divided into multiple training batches according to a preset number. In each iteration, the model parameters are updated using mini-batch stochastic gradient descent, training is performed using multiple training batches, and the training loss for each sample and each batch is calculated using the loss function. Updating the model parameters using mini-batch stochastic gradient descent, and selecting a fixed number of training set samples to form a batch for training in each iteration, can balance training efficiency and parameter update stability.

[0048] It should be noted that after each iteration, the training samples in the validation set need to be input into the updated convolutional neural network to obtain the recognition result of each training sample. Then, the loss value of the validation set is calculated using the loss function corresponding to the training batch. The recognition result of each training sample is compared with the actual situation of that training sample. If they match, it means that the training sample is recognized accurately; otherwise, it means that the training sample is recognized inaccurately. The ratio of the number of accurately recognized training samples to the total number of training samples in the validation set is used as the recognition accuracy of the convolutional neural network model obtained in this iteration on the validation set.

[0049] It should be noted that the convergence criterion is that the improvement in the recognition accuracy of the validation set is lower than a preset value for multiple consecutive iterations, and the loss value of the validation set maintains a downward trend or stabilizes below a preset threshold for multiple consecutive iterations. If the updated convolutional neural network model does not meet the above convergence conditions, the learning rate is adjusted using a learning rate decay strategy. The initial learning rate of the convolutional neural network model can be set to 0.001, and the learning rate decay strategy can be to reduce the learning rate to 0.9 times the learning rate of the previous 5 iterations every 5 iterations. While adjusting the learning rate, the model weights and bias parameters also need to be updated before continuing the iteration. If the updated convolutional neural network model meets the above convergence conditions, the model recognition performance needs to be finally evaluated using a test set. In addition, when the model does not converge, an adaptive momentum estimation optimization algorithm can be used to update the model parameters. This algorithm can adaptively adjust the learning rate and improve the convergence speed of model training.

[0050] In some embodiments, when performing a final evaluation test on the converged convolutional neural network model using a test set, it is necessary to calculate the recognition accuracy of the test set. When the recognition accuracy of the test set is not lower than a preset standard, the iteratively updated convolutional neural network model is determined as a pre-trained convolutional neural network model. When the accuracy of the test set is lower than the preset standard, it is necessary to adjust the training set of the model, or adjust the training method and hyperparameters of the model.

[0051] In one possible implementation, the loss function corresponding to the training batch is:

[0052] in, This refers to the loss value for each training batch when training is performed in batches. The number of training samples in each batch. The first in each batch training samples, The total number of categories to be separated. For the first The weighting coefficients of the class separation categories. For the first The training sample belongs to the first... The true label of the class, For the first The training sample belongs to the i-th training sample after model output. The predicted probability value of the class; The loss function corresponding to the training samples is:

[0053]

[0054] in, The training loss value for each training sample. For balance coefficient, The detection loss is for asphalt mixture images in the training samples. Image of asphalt mixture. The detection loss is the standard image from the training samples. For standard images, This represents the loss from selective self-supervised regression.

[0055] By combining image preprocessing with an adaptive low-light image enhancement algorithm, the quality of low-light asphalt mixture images was effectively optimized, significantly improving the recognition accuracy of segregation features in asphalt mixture images. The convolutional neural network model employed, through a feature extraction module with multi-size convolutional kernels, improved the stability of feature distribution and nonlinear expression capabilities, enhancing the representation effect on complex target features. Combined with a multi-level pooling structure, ordered dimensionality reduction of the feature map was achieved, strengthening key local features while fully preserving overall feature information, effectively improving the model's feature selection capability and computational efficiency. Simultaneously, during the model training phase, by combining a composite loss function with batch and sample dimensions, along with a learning rate decay strategy and phased dataset validation testing, the training efficiency and convergence effect of the model were effectively improved, enhancing its generalization ability. Ultimately, it can accurately capture the feature differences of different segregation degrees in low-light images, significantly improving the accuracy and detection efficiency of asphalt mixture segregation identification under low-light conditions, and better meeting the rapid segregation detection needs in actual construction scenarios such as nighttime and low-light conditions.

[0056] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0057] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0058] Figure 4 A schematic diagram of the asphalt mixture segregation identification device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 4 As shown, the asphalt mixture segregation identification device 4 includes: The preprocessing module 41 is used to perform image preprocessing on the acquired asphalt mixture image to be identified, so as to obtain the preprocessed image to be identified. Enhancement module 42 is used to perform image enhancement processing on the preprocessed image to be identified based on a preset low-light image enhancement algorithm to obtain a standard image of asphalt mixture; The recognition module 43 is used to input the standard image into the pre-trained convolutional neural network model to obtain the segregation recognition result of the asphalt mixture; wherein, the pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module and an output module; wherein, the feature extraction module includes convolutional units with convolutional kernels of different sizes; the pooling module includes a multi-level pooling structure.

[0059] In one possible implementation, the enhancement module 42 is specifically used to: determine an adaptive image enhancement formula according to a preset low-light image enhancement algorithm; and perform image enhancement on the pixel grayscale values ​​at each position of the preprocessed image to be identified based on the adaptive image enhancement formula to obtain a standard image of asphalt mixture.

[0060] In one possible implementation, the recognition module 43 is specifically used for: inputting a standard image into a general feature extraction module to obtain a general visual feature map of the asphalt mixture; inputting the general visual feature map into a feature extraction module to obtain a multi-scale feature map of the asphalt mixture; inputting the multi-scale feature map into a pooling module to obtain a global average pooling feature vector of the asphalt mixture; inputting the global average pooling feature vector into a fully connected module to obtain a dimensionality-reduced feature vector of the asphalt mixture; and inputting the dimensionality-reduced feature vector into an output module, using the softmax function in the output module to predict the segregation probability of the dimensionality-reduced feature vector to obtain the segregation recognition result of the asphalt mixture.

[0061] In one possible implementation, the recognition module 43 is further configured to: input a general visual feature map into a first convolutional unit to obtain a fine-grained texture feature map of the asphalt mixture; input the fine-grained texture feature map into a second convolutional unit to obtain a segregation region contour feature map of the asphalt mixture; and input the segregation region contour feature map into a third convolutional unit to obtain a multi-scale feature map of the asphalt mixture.

[0062] In one possible implementation, the recognition module 43 is further configured to: input a fine-grained texture feature map into a first pooling layer to obtain a first dimensionality-reduced feature map of the asphalt mixture; input a segregation region contour feature map into a second pooling layer to obtain a second dimensionality-reduced feature map of the asphalt mixture; input a multi-scale feature map into a third pooling layer to obtain an average feature map of the asphalt mixture; and input the multi-scale feature map, the first dimensionality-reduced feature map, the second dimensionality-reduced feature map, and the average feature map into a global average pooling layer to obtain a global average pooling feature vector of the asphalt mixture.

[0063] In one possible implementation, the recognition module 43 is further configured to: input the global average pooling feature vector into the first hidden layer, perform depth mapping on the global average pooling feature vector based on the neurons in the first hidden layer to obtain the mapping feature vector of the asphalt mixture; input the mapping feature vector into the second hidden layer, and perform feature dimension compression on the mapping feature vector based on the neurons in the second hidden layer to obtain the dimension-reduced feature vector of the asphalt mixture.

[0064] In one possible implementation, the training process of the pre-trained convolutional neural network model includes: performing image preprocessing and image enhancement processing on multiple normal asphalt mixture images and multiple asphalt mixture images with different degrees of segregation, respectively, to obtain a standard image corresponding to each asphalt mixture image; using each asphalt mixture image and its corresponding standard image as a training sample, and dividing all training samples into a training set, a validation set, and a test set according to a preset ratio; dividing all shuffled training samples from the training set into multiple training batches according to a preset number; iteratively updating the model parameters of the convolutional neural network model based on the training samples and loss function in each training batch; and after each iteration, The training samples from the validation set are input into the iteratively updated convolutional neural network model to obtain the recognition result of each training sample. Based on the recognition result of each training sample and the model convergence criterion, it is determined whether the iteratively updated convolutional neural network model has converged. If the iteratively updated convolutional neural network model does not converge, the learning rate in each subsequent iteration of training is adjusted based on the learning rate decay strategy, and the model parameters of the convolutional neural network model are iteratively updated based on the adjusted learning rate, the training samples in each training batch, and the loss function. If the iteratively updated convolutional neural network model converges, the iteratively updated convolutional neural network model is tested based on the training samples in the test set to obtain the pre-trained convolutional neural network model.

[0065] In one possible implementation, the loss function includes a loss function corresponding to the training batch and a loss function corresponding to each training sample; the loss function corresponding to the training batch is:

[0066] in, This refers to the loss value for each training batch when training is performed in batches. The number of training samples in each batch. The first in each batch training samples, The total number of categories to be separated. For the first The weighting coefficients of the class separation categories. For the first The training sample belongs to the first... The true label of the class, For the first The training sample belongs to the i-th training sample after model output. The predicted probability value of the class; The loss function corresponding to the training samples is:

[0067]

[0068] in, The training loss value for each training sample. For balance coefficient, The detection loss is for asphalt mixture images in the training samples. Image of asphalt mixture. The detection loss is the standard image from the training samples. For standard images, This represents the loss from selective self-supervised regression.

[0069] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 5 As shown, the electronic device 5 of this embodiment includes a processor 50 and a memory 51. The memory 51 stores a computer program 52. When the processor 50 executes the computer program 52, it implements the steps in the various method embodiments described above. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each module / unit in the various device embodiments described above.

[0070] For example, computer program 52 may be divided into one or more modules / units, which are stored in memory 51 and executed by processor 50 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 52 in electronic device 5.

[0071] Electronic device 5 may include, but is not limited to, processor 50 and memory 51. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 5 and does not constitute a limitation on electronic device 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 5 may also include input / output devices, network access devices, buses, etc.

[0072] The processor 50 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0073] The memory 51 can be an internal storage unit of the electronic device 5, such as a hard disk or RAM. The memory 51 can also be an external storage device of the electronic device 5, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 51 can include both internal and external storage units of the electronic device 5. The memory 51 is used to store the computer program 52 and other programs and data required by the electronic device 5. The memory 51 can also be used to temporarily store data that has been output or will be output.

[0074] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0075] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0076] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying segregation in asphalt mixtures, characterized in that, include: The collected images of asphalt mixtures to be identified are preprocessed to obtain the preprocessed images to be identified. Based on a preset low-light image enhancement algorithm, the preprocessed image to be identified is subjected to image enhancement processing to obtain a standard image of the asphalt mixture; The standard image is input into a pre-trained convolutional neural network model to obtain the segregation identification result of the asphalt mixture; wherein, the pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module; wherein, the feature extraction module includes convolutional units with convolutional kernels of different sizes; the pooling module includes a multi-level pooling structure.

2. The method for identifying asphalt mixture segregation according to claim 1, characterized in that, The preset low-light image enhancement algorithm performs image enhancement processing on the preprocessed image to be identified to obtain a standard image of the asphalt mixture, including: Based on the preset low-light image enhancement algorithm, determine the adaptive image enhancement formula; Based on the adaptive image enhancement formula, the pixel grayscale values ​​at each position of the preprocessed image to be identified are enhanced to obtain the standard image of the asphalt mixture.

3. The method for identifying asphalt mixture segregation according to claim 1, characterized in that, The step of inputting the standard image into a pre-trained convolutional neural network model to obtain the segregation identification result of the asphalt mixture includes: The standard image is input into the general feature extraction module to obtain the general visual feature map of the asphalt mixture; The general visual feature map is input into the feature extraction module to obtain the multi-scale feature map of the asphalt mixture; The multi-scale feature map is input into the pooling module to obtain the global average pooling feature vector of the asphalt mixture; The global average pooling feature vector is input into the fully connected module to obtain the dimensionality-reduced feature vector of the asphalt mixture; The reduced feature vector is input into the output module, and the segregation probability is predicted by the softmax function in the output module to obtain the segregation identification result of the asphalt mixture.

4. The method for identifying asphalt mixture segregation according to claim 3, characterized in that, The feature extraction module includes a first convolutional unit, a second convolutional unit, and a third convolutional unit; the convolutional kernels of the first convolutional unit, the second convolutional unit, and the third convolutional unit have different sizes; the step of inputting the general visual feature map into the feature extraction module to obtain a multi-scale feature map of the asphalt mixture includes: The general visual feature map is input into the first convolutional unit to obtain the fine-grained texture feature map of the asphalt mixture; The fine-grained texture feature map is input into the second convolutional unit to obtain the contour feature map of the segregation region of the asphalt mixture; The contour feature map of the segregated region is input into the third convolutional unit to obtain the multi-scale feature map of the asphalt mixture.

5. The method for identifying asphalt mixture segregation according to claim 4, characterized in that, The multi-level pooling structure includes a first pooling layer, a second pooling layer, a third pooling layer, and a global average pooling layer; the step of inputting the multi-scale feature map into the pooling module to obtain the global average pooling feature vector of the asphalt mixture includes: The fine-grained texture feature map is input into the first pooling layer to obtain the first dimension-reduced feature map of the asphalt mixture; The contour feature map of the segregation region is input into the second pooling layer to obtain the second dimension-reduced feature map of the asphalt mixture; The multi-scale feature map is input into the third pooling layer to obtain the average feature map of the asphalt mixture; The multi-scale feature map, the first dimensionality-reduced feature map, the second dimensionality-reduced feature map, and the average feature map are input into the global average pooling layer to obtain the global average pooling feature vector of the asphalt mixture.

6. The method for identifying asphalt mixture segregation according to claim 5, characterized in that, The fully connected module includes a first hidden layer and a second hidden layer; the number of neurons in the first hidden layer is greater than the number of neurons in the second hidden layer; the step of inputting the global average pooling feature vector into the fully connected module to obtain the dimensionality-reduced feature vector of the asphalt mixture includes: The global average pooling feature vector is input into the first hidden layer, and the global average pooling feature vector is depth-mapped based on the neurons in the first hidden layer to obtain the mapping feature vector of the asphalt mixture. The mapped feature vector is input into the second hidden layer, and the feature dimension of the mapped feature vector is compressed based on the neurons in the second hidden layer to obtain the dimension-reduced feature vector of the asphalt mixture.

7. The method for identifying asphalt mixture segregation according to any one of claims 1-6, characterized in that, The training process of the pre-trained convolutional neural network model includes: For multiple normal asphalt mixture images and multiple asphalt mixture images with different degrees of segregation, image preprocessing and image enhancement processing are performed respectively to obtain the standard image corresponding to each asphalt mixture image; Each asphalt mixture image and its corresponding standard image are used as a training sample, and all training samples are divided into a training set, a validation set, and a test set according to a preset ratio. The training set is divided into multiple training batches according to a preset number of shuffled training samples. The model parameters of the convolutional neural network model are iteratively updated based on the training samples and loss function in each training batch. After each iteration, the training samples in the validation set are input into the iteratively updated convolutional neural network model to obtain the recognition result of each training sample. Based on the recognition result of each training sample and the model convergence judgment criteria, it is determined whether the iteratively updated convolutional neural network model has converged. If the convolutional neural network model does not converge after iterative updates, the learning rate in each subsequent iteration of training is adjusted based on the learning rate decay strategy, and the model parameters of the convolutional neural network model are iteratively updated based on the adjusted learning rate, the training samples in each training batch, and the loss function. If the iteratively updated convolutional neural network model converges, then the iteratively updated convolutional neural network model is tested based on the training samples in the test set to obtain the pre-trained convolutional neural network model.

8. The method for identifying asphalt mixture segregation according to claim 7, characterized in that, The loss function includes the loss function corresponding to the training batch and the loss function corresponding to each training sample; The loss function corresponding to the training batch is: in, This refers to the loss value for each training batch when training is performed in batches. The number of training samples in each batch. The first in each batch training samples, The total number of categories to be separated. For the first The weighting coefficients of the class separation categories. For the first The training sample belongs to the first... The true label of the class, For the first The training sample belongs to the i-th training sample after model output. The predicted probability value of the class; The loss function corresponding to the training samples is: in, The training loss value for each training sample. For balance coefficient, The detection loss is the asphalt mixture image in the training samples. Image of asphalt mixture. The detection loss is the standard image in the training samples. For standard images, This represents the loss from selective self-supervised regression.

9. A device for identifying asphalt mixture segregation, characterized in that, include: The preprocessing module is used to preprocess the acquired asphalt mixture images to be identified, and obtain the preprocessed images to be identified. An enhancement module is used to perform image enhancement processing on the preprocessed image to be identified based on a preset low-light image enhancement algorithm to obtain a standard image of the asphalt mixture. The recognition module is used to input the standard image into a pre-trained convolutional neural network model to obtain the segregation recognition result of the asphalt mixture; wherein, the pre-trained convolutional neural network model includes a general feature extraction module, a feature extraction module, a pooling module, a fully connected module, and an output module; wherein, the feature extraction module includes convolutional units with convolutional kernels of different sizes; the pooling module includes a multi-level pooling structure.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images