Over-fitting Problem Research Articles

Hybrid quantum neural network based on weight remapping and its applications

Abstract In order to explore the possibility of cross-fertilization between quantum computing and neural networks, and to analyse the effects of multiple weight remapping functions on the model performance, this article proposes two hybrid models based on weight remapping: the hybrid quantum neural network (HQNN-WR) and the hybrid quantum convolutional neural network (HQCNN-WR). The HQNN-WR model uses a two-layer fully connected network to process the input features, performs feature transformation by applying multiple weight remapping functions, and subsequently passes the processed features to the quantum neural network for computation, and finally outputs the classification results. The experimental results show that the model significantly improves the classification accuracy on Iris, Wine and Breast datasets and the convergence speed is faster. The HQCNN-WR model integrates convolutional layers, pooling layers, and fully connected layers, and prevents over-fitting through a dropout layer, and exhibits excellent performance in binary classification tasks on MNIST and KMNIST datasets. The model effectively mitigates the over-fitting problem on small sample datasets and enhances the robustness and generalization ability of the model while improving the digit recognition accuracy. By comparing different models, this article also demonstrates their significant effects on the performance of hybrid quantum neural networks, providing a new theoretical basis and experimental support for the optimization and application of quantum machine learning methods.

A comparison of representations in grammar-guided genetic programming in the context of glucose prediction in people with diabetes

The representation of individuals in Genetic Programming (GP) has a large impact on the evolutionary process. In previous work, we investigated the evolutionary process of three Grammar-Guided GP (GGGP) methods, Context-Free Grammars GP (CFG-GP), Grammatical Evolution (GE) and Structured Grammatical Evolution (SGE), in the context of the complex, real-world problem of predicting the glucose level of people with diabetes two hours ahead of time. We concluded that representation choice is more impactful with a higher maximum depth, and that CFG-GP better explores the search space for deeper trees, achieving better results. Furthermore, we find that CFG-GP relies more on feature construction, whereas GE and SGE rely more on feature selection. Additionally, we altered the GGGP methods in two ways: using ϵ-lexicase selection, which solved the overfitting problem of CFG-GP and helps it to adapt to patients with high glucose variability; and with a penalization of complex trees, to create more interpretable trees. Combining ϵ-lexicase selection with CFG-GP performed best. In this work, we extend on the previous work and evaluated the impact of initialization methods in the quality of solutions. We found that they have no significant impact, even when the change of representation has.

Application of convolutional neural networks in image classification and applications of improved convolutional neural networks

Abstract. This paper reviews the application and improvement of convolutional neural networks (CNNs) in image classification. Firstly, a shallow CNN for interstitial lung disease image classification is presented. This model suppresses overfitting through a unique network architecture and optimisation algorithm. Next, the improved VGG16 architecture and MIDNet18 model are discussed and their superior performance in brain tumour image classification is demonstrated. Subsequently, a CNN-CapsNet model for cervical cancer image classification and its improvement are presented and the customised model is compared with the conventional VGG-16 CNN architecture in the paper. Next, the application of sparse convolutional kernels and hybrid sparse convolutional kernels (HDCs) in solving the problem of computational resource consumption is presented. Subsequently, methods for solving the problem of limited training data through transfer learning and network data augmentation techniques are discussed, as well as GAN-generated datasets for solving the overfitting problem. Finally, the effect of degraded images on the classification effectiveness of CNNs is explored. The results show that the improved CNN architecture and algorithms have significant effects in solving the problems of overfitting and computational resource consumption, and can significantly improve the accuracy and efficiency of image classification. And degraded images do adversely affect the accuracy of CNN for image classification.

HTCSigNet: A Hybrid Transformer and Convolution Signature Network for offline signature verification

For Offline Handwritten Signature Verification (OHSV) tasks, traditional Convolutional Neural Networks (CNNs) and transformers are hard to individually capture global and local features from signatures, and single-depth models often suffer from overfitting and poor generalization problems. To overcome those difficulties, in this paper, a novel Hybrid Transformer and Convolution Signature Network (HTCSigNet) is proposed to capture multi-scale features from signatures. Specifically, the HTCSigNet is an innovative framework that consists of two parts: transformer and CNN-based blocks which are used to respectively extract global and local features from signatures. The CNN-based block comprises a Space-to-depth Convolution (SPD-Conv) module which improves the feature learning capability by precisely focusing on signature strokes, a Spatial and Channel Reconstruction Convolution (SCConv) module which enhances model generalization by focusing on more distinctive micro-deformation features while reducing attention to common features, and convolution module that extracts the shape, morphology of specific strokes, and other local features from signatures. In the transformer-based block, there is a Vision Transformer (ViT) which is used to extract overall shape, layout, general direction, and other global features from signatures. After the feature learning stage, Writer-Dependent (WD) and Writer-Independent (WI) verification systems are constructed to evaluate the performance of the proposed HTCSigNet. Extensive experiments on four public signature datasets, GPDSsynthetic, CEDAR, UTSig, and BHSig260 (Bengali and Hindi) demonstrate that the proposed HTCSigNet learns discriminative representations between genuine and skilled forged signatures and achieves state-of-the-art or competitive performance compared with advanced verification systems. Furthermore, the proposed HTCSigNet is easy to transfer to different language datasets in OHSV tasks.22The code is available at https://github.com/copycpp/HTCSigNet-Master.

Financial Time Series Forecasting Based on Long-Short Term Memory, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Batch Normalization

Long-short term memory (LSTM) is a state-of-art and widely used model to forecast financial time series. However, primitive LSTM networks do not perform well due to over-fitting problems of the deep learning model and nonlinear and non-stationary characteristics of financial time series data. In addition, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is an outstanding data frequency decomposition technique that can decompose original time series into several intrinsic mode functions and a residue. Thus, this paper proposed a novel hybrid network CEEMDAN-LSTM-BN based on LSTM. Specifically, to avoid over-fitting, the modified LSTM-BN network consists of two LSTM layers, two Batch Normalization (BN) layers following each LSTM layer, and a dropout layer. Each of the intrinsic mode functions and the residue would be processed by CEEMDAN-LSTM-BN and the final prediction results are obtained by reconstructing each predictive series. The advantages of the proposed CEEMDANLSTM-BN networks are verified by comparing them to primitive LSTM, other hybrid models, and some famous machine learning models. Moreover, the robustness of the networks is assessed by numerical experiments on different stock indices datasets.

Predictive slope stability early warning model based on CatBoost

A model for predicting slope stability is developed using Categorical Boosting (CatBoost), which incorporates 6 slope features to characterize the state of slope stability. The model is trained using a symmetric tree as the base model, utilizing ordered boosting to replace gradient estimation, which enhances prediction accuracy. Comparative models including Support Vector Machine (SVM), Light Gradient Boosting Machine (LGBM), Random Forest (RF), and Logistic Regression (LR) were introduced. Five performance evaluation metrics are utilized to assess the predictive capabilities of the CatBoost model. Based on CatBoost model, the predicted probability of slope instability is calculated, and the early warning model of slope instability is further established. The results suggest that the CatBoost model demonstrates a 6.25% disparity in accuracy between the training and testing sets, achieving a precision of 100% and an Area Under Curve (AUC) value of 0.95. This indicates a high level of predictive accuracy and robust ordering capabilities, effectively mitigating the problem of overfitting. The slope instability warning model offers reasonable classifications for warning levels, providing valuable insights for both research and practical applications in the prediction of slope stability and instability warning.

HSMix: Hard and soft mixing data augmentation for medical image segmentation

Due to the high cost of annotation or the rarity of some diseases, medical image segmentation is often limited by data scarcity and the resulting overfitting problem. Self-supervised learning and semi-supervised learning can mitigate the data scarcity challenge to some extent. However, both of these paradigms are complex and require either hand-crafted pretexts or well-defined pseudo-labels. In contrast, data augmentation represents a relatively simple and straightforward approach to addressing data scarcity issues. It has led to significant improvements in image recognition tasks. However, the effectiveness of local image editing augmentation techniques in the context of segmentation has been less explored. Additionally, traditional data augmentation methods for local image editing augmentation methods generally utilize square regions, which cause a loss of contour information.We propose HSMix, a novel approach to local image editing data augmentation involving hard and soft mixing for medical semantic segmentation. In our approach, a hard-augmented image is created by combining homogeneous regions (superpixels) from two source images. A soft mixing method further adjusts the brightness of these composed regions with brightness mixing based on locally aggregated pixel-wise saliency coefficients. The ground-truth segmentation masks of the two source images undergo the same mixing operations to generate the associated masks for the augmented images.Our method fully exploits both the prior contour and saliency information, thus preserving local semantic information in the augmented images while enriching the augmentation space with more diversity. Our method is a plug-and-play solution that is model agnostic and applicable to a range of medical imaging modalities. Extensive experimental evidence has demonstrated its effectiveness in a variety of medical segmentation tasks. The source code is available in https://github.com/DanielaPlusPlus/HSMix.

Machine learning-driven polygenic risk scores for bipolar disorder, depression, and panic disorder

Polygenic Risk Score (PRS) is a computational tech- nique that uses various genomic data to simultaneously analyze an individ- ual’s genetic risk for particular illnesses or traits. However, the traditional PRS computation has a few weaknesses, including its limited capacity to account for just a portion of trait variance, susceptibility to overfitting, and insufficient ability to discriminate among the larger population. Machine Learning (ML) methods offer a promising alternative to the traditional method by avoiding the problem of overfitting and improving accuracy. This study aims to develop an ML model for improved PRS calculation. We used the summary statistics for three mentals diseases, bipolar, depression, and panic disorder, from the Psychiatric Genomics Consortium (PGC) as a disease reference. We also obtained actual genotype data of individuals from OpenSNP, which includes both case and control samples. This data is used for predicting scores. The suggested approach, called Polygenic Risk Score Neural Network (PRSNN), calculates the PRS using weight vectors that estimate the relevance of each single nucleotide polymorphism (SNP) with a particular phenotype by deep learning model as an alternative to the traditional method. This study aims to develop a machine learning model, called PRSNN, for improved calculation of Polygenic Risk Scores (PRS). The PRSNN method outperforms the conventional method in identifying individuals at risk of mental disease. A novel deep-learning approach, named as PRSNN, is proposed for generating PRSs. The results demonstrate that it outperforms the traditional method of computing PRS for complex diseases. Further upgrades for this tool are required to overcome the current limitations, including lack of validation with external data from different ancestries, which may limit the applicability of the PRSNN method across diverse populations, and the small sample size, which may affect the results.

Styled and characteristic Peking opera facial makeup synthesis with co-training and transfer conditional styleGAN2

Against the backdrop of the deep integration of culture and technology, research and practice in digitization of intangible cultural heritage has continued to deepen. However, due to the lack of data and training, it is still very difficult to apply artificial intelligence to the field of cultural heritage protection. This article integrates image generation technology into the digital protection of Peking opera facial makeup, using a self-built Peking opera facial makeup dataset. Based on the StyleGAN2 network, we propose a style generative cooperative training network Co-StyleGAN2, which integrates the adaptive data augmentation (ADA) to alleviate the problem of discriminator overfitting and introduces the idea of cooperative training to stabilize the training process. We design a Peking opera facial makeup image transform conditional generation network TC-StyleGAN2 which is transferred from unconditional generation network. The weights of the unconditional pre-training model are fixed, and an adaptive filtering modulation module is added to modulate the category parameters to complete the conversion from unconditional to conditional StyleGAN2 to deal with the training difficulty of conditional GANs on limited data, which suffer from severe mode collapse. The experimental results show that the proposed training strategy is better than the comparison algorithms, and the image generation quality and diversity have been improved.

CGAN Facilitated Data Augmentation of Voice and Speech Parameters for Detecting Parkinson’s Disease in the Prodromal Phase

Parkinson’s disease is a multi-faceted disease affecting the brain. The enormity of its recent rise is quite alarming. This calls for intense research to diagnose early to hasten the progress of diagnosis. Voice distortion is considered an early precursor for Parkinson’s disease. Though several studies in Machine Learning using voice parameters have provided useful information, none of them have been successful in evolving an efficient and generalized model to detect it. Deep Learning techniques were applied to improve the performance of the model but its major limitation was the size of the dataset. Hence, a need arose to extend the dataset using an appropriate data augmentation method. At this juncture, the conditional generative adversarial network (CGAN) proved to be a useful technique because of its innate feature for generating synthetic data from input noise. The RNN-LSTM classifier could achieve a training accuracy of 87.32%, testing accuracy of 86.3%, training precision of 87.92 %, and testing precision of 89.94%. The results of the experimental study are compared with other state-of-the-art methods. This technique succeeded in reducing the problem of over-fitting and could elevate the performance of the RNN-LSTM classifier in the prediction of Parkinson’s disease.

Analysis of health recommendations using longitudinal quality of life data: QoL@TbA - A transformer-based approach.

Objective: Health recommendation systems suggest behavioral modifications to improve quality of life. However, current approaches do not facilitate the generation or examination of such recommendations considering the multifeature longitudinal evolution of behaviors. This paper proposes the use of a deep learning transformer-based model that allows the analysis of recommendations for behavior changes. Methods: We adapted a prediction approach, namely Behavior Sequence Transformer (BST), which analyzes temporal human routines and patterns, generating inductive outcomes. The evaluation relied on a case study that employed the behavioral history and profile of the English Longitudinal Study of Ageing (ELSA) participants (n = 2682), predicting their psychological mood (normal, pre-depressed, depressed) according to input recommendations for behavioral changes. Root mean squared error (RMSE) and learning curves were used to track the recommendation accuracy evolution and possible overfitting problems. Results: Experiments demonstrated lower RMSE values for the multifeature model (0.28/0.03) when compared to its single-feature versions (marital status, 0.59/0.001), (high pressure, 0.357/0.04), (diabetes, 0.36/0.01), (sleep quality, 0.57/0.02), (level of physical activity, 0.57/0.01). Conclusions: The results demonstrate the architecture's capability to analyze multifeatured longitudinal data, supporting the generation of suggestions for concurrent modifications across multiple input features. Moreover, these suggestions align with findings in specialized literature.

The Application of Deep Learning in Sports Training

In response to the problems of overfitting, susceptibility to interference information, and insufficient feature expression ability in existing deep learning methods for sports action recognition, the author proposes a deep learning sports action recognition method that integrates attention mechanism. This method proposes a video data augmentation algorithm in data preprocessing to reduce the risk of model overfitting. Then, during the video frame sampling process, the existing sampling algorithms are improved to effectively suppress the influence of interference information. In the special section, the network residual consolidation is proposed to improve the feature extraction capacity of the structure. The Long Term Time Transform (LSTM) network is used to solve the problem of the global correlation of the spatial correlation, and the classification algorithm is achieved by Softmax. and the classification algorithm is proposed. The experimental results show that the recognition rates of this method on UCFYouTube, KTH, and HMDB-51 data are 96.73%, 98.07%, and 64.82%, respectively.

Long-term Evaluation of Machine Learning Based Methods for Air Emission Monitoring.

Machine learning (ML) techniques have been researched and used in various environmental monitoring applications. Few studies have reported the long-term evaluation of such applications. Discussions regarding the risks and regulatory frameworks of ML applications in environmental monitoring have been rare. We monitored the performance of six predictive models developed using ML and statistical methods for 28 months. The six models used to predict NOx emissions were developed using six different algorithms. The model developed with a moderate complexity algorithm, adaptive boosting, had the best performance in long-term monitoring, with a root mean square error (RMSE) of 0.48 kg/hr in the 28-month monitoring period, and passed two of the three relative accuracy test audits. High complexity models based on gradient boosting and neural network algorithms had the best training performance, with a minimum RMSE of 0.23 kg/hr and 0.26 kg/hr, but also had the worst RMSE scores, of 0.51 kg/hr and 0.57 kg/hr, during the monitoring period. In addition, all six models failed all three relative accuracy test audits. The following problems were observed: (1) Complex ML models tended to have overfitting problems, thus indicating the importance of the trade-off between model accuracy and complexity. (2) Model input sensor drift or out of high-frequency ranges from the training data resulted in inaccurate predictions or an accuracy lower than the minimum allowed by regulators. (3) Existing regulatory frameworks must be modernized to keep pace with current machine learning practices. Some statistical tests are unsuitable for applications developed by using ML methods.

PGNN-Net: Parallel Graph Neural Networks for Hyperspectral Image Classification Using Multiple Spatial-Spectral Features

Hyperspectral image (HSI) shows great potential for application in remote sensing due to its rich spectral information and fine spatial resolution. However, the high dimensionality, nonlinearity, and complex relationship between spectral and spatial features of HSI pose challenges to its accurate classification. Traditional convolutional neural network (CNN)-based methods suffer from detail loss in feature extraction; Transformer-based methods rely too much on the quantity and quality of HSI; and graph neural network (GNN)-based methods provide a new impetus for HSI classification by virtue of their excellent ability to handle irregular data. To address these challenges and take advantage of GNN, we propose a network of parallel GNNs called PGNN-Net. The network first extracts the key spatial-spectral features of HSI using principal component analysis, followed by preprocessing to obtain two primary features and a normalized adjacency matrix. Then, a parallel architecture is constructed using improved GCN and ChebNet to extract local and global spatial-spectral features, respectively. Finally, the discriminative features obtained through the fusion strategy are input into the classifier to obtain the classification results. In addition, to alleviate the over-fitting problem, the label smoothing technique is embedded in the cross-entropy loss function. The experimental results show that the average overall accuracy obtained by our method on Indian Pines, Kennedy Space Center, Pavia University Scene, and Botswana reaches 97.35%, 99.40%, 99.64%, and 98.46%, respectively, which are better compared to some state-of-the-art methods.

Optimization of Classification Algorithms Performance with k-Fold Cross Validation

Supervised learning is a predictive method used to make predictions or classifications. Supervised learning algorithms work by building a model using training data that includes both independent and dependent variables. Several methods for building classification include Logistic Regression, Naive Bayes, K-Nearest Neighbor (KNN), decision tree, etc. The lack of capacity of a classification algorithm to generalize certain data can be associated with the problem of overfitting or underfitting. K-fold cross-validation is a method that can help avoid overfitting or underfitting and produce a algorithm with good performance on new data. This study will test the Naive Bayes, K-Nearest Neighbor (KNN), Classification and Regression Tree (CART), and Logistic Regression methods with k-fold cross-validation on two different datasets. The values of k set for cross-validation are 2, 3, 5, 7, and 10. The analysis results concluded that each classification algorithm performed best at 10-fold cross-validation. In DATA 1, the Naive Bayes algorithm has the highest average accuracy of 0.67 (67%) and the error rate is 0.33 (33%), followed by the CART algorithm, KNN, and finally logistic regression. While DATA 2, the KNN algorithm has the highest average accuracy of 0.66 (66%) and an error rate of 0.34 (34%), followed by the CART algorithm, Naive Bayes, and finally logistic regressionbut can be a reference if you want to predict the growth direction of the accommodation and food service activities sector.

Deep learning algorithms for addressing overfitting and biological realism in tree taper and volume predictions

This study addresses the challenges of overfitting and maintaining biological realism in deep learning algorithms (DLAs), for predicting individual tree taper using stem diameters outside bark (DOB) and total tree volume (TTV). To this end, DLAs were trained using two different approaches: a “hyperparameter-optimized DLA”, which customizes specific hyperparameters such as learning rate and momentum rate, and a “regularization-optimized DLA”, which incorporates optimization techniques like early stopping with root mean square error, L1 and L2 regularization, and dropout. Although obtaining the deterioration in predictive capabilities statistics from the taring dataset to the validation dataset by standard DLA with adaptive learning processes without customizing the hyperparameters and regularization parameters, the hyperparameter-optimized DLA with a momentum of 0.8, and a 7 # hidden layer for the TTV and regularization-optimized DLA with a dropout ratio of 0.000001, a 3 # hidden layer for the DOB demonstrated comparable predictive capabilities statistics across both training and validation datasets with generating biologically plausible predictions. Our results support that these hyperparameter-optimized and regularization-optimized DLAs, by improving the “black-box” nature of artificial intelligence, offer significant potential for enhanced interpretability and performance by improving the problem of overfitting and the violations biological realism in forest biometrics applications.

Adaptive Neighbors Graph Learning for Large-Scale Data Clustering using Vector Quantization and Self-Regularization

In traditional adaptive neighbors graph learning (ANGL)-based clustering, the time complexity is more than O(n2), where n is the number of data points, which is not scalable for large-scale data problems in real applications. Subsequently, ANGL adds a balance regularization to its objective function to avoid the sparse over-fitting problem in the learned similarity graph matrix. Still, the regularization may leads to many weak connections between data points in different clusters. To address these problems, we propose a new fast clustering method, namely, Adaptive Neighbors Graph Learning for Large-Scale Data Clustering using Vector Quantization and Self-Regularization (ANGL-LDC), to perform vector quantization (VQ) on original data and feed the obtained VQ data as the input in the n×n similarity graph matrix learning. Hence, the n×n similarity graph matrix learning problem is simplified to weighted m×m(m≪n) graph learning problem, where m is the number of distinct points and weight is the duplicate times of distinct points in VQ data. Consequently, the time complexity of ANGL-LDC is much lower than that of ANGL. At the same time, we propose a new ANGL objective function with a graph connection self-regularization mechanism, where the ANGL-LDC objective function will get an infinity value if the value of one graph connection is equal to 1. Therefore, ANGL-LDC naturally avoids obtaining the sparse over-fitting problem since we need to minimize the value of ANGL-LDC’s objective function. Experimental results on synthetic and real-world datasets demonstrate the scalability and effectiveness of ANGL-LDC.

Automated medical diagnosis of dementia through fine-tuned EfficientNet

Automated models using deep learning are more extensively used in medical imaging in the last decade. The present study contributes to the diagnosis of dementia using MRI. Dementia is a syndrome that deteriorates the cognitive function of the brain. The disease has no cure, till now, except for the prior diagnosis. The present study aims for classifying the MRI scans of two datasets OASIS and ADNI into 2 categories: binary and multi-classification. To achieve the objective, the EfficientNetB0 architecture of deep learning is fine-tuned by adding three dense layers on the top of the network. The swish activation function is used in the inner dense layers added. The dropout and batch normalization layers are also added for dealing with the problem of overfitting. This architecture offers high accuracy and high efficiency compared to other pre-trained networks. The model is assessed on various performance measures and outperformed the state of art techniques. For the OASIS dataset, the best testing accuracy for binary classification is 93.10% with a 0.01 learning rate. The sensitivity is 95.93%, specificity is 90.08%, false-negative rate is 4.07, the false-positive rate is 9.92 and the F1-score is 93.48%. The best testing accuracy of multi-classification is 84.50% with a 0.001 learning rate. For the ADNI dataset, the best testing accuracy is 96.08% with a learning rate of 0.001. The sensitivity is 94.74%, specificity is 99.32%, false-negative rate is 5.26, the false-positive rate is 0.68 and the f1-score is 97.16%. The best testing accuracy of multi-classification is 98.10 with a 0.01 learning rate. The proposed model can be utilized for developing an automated framework to help medical services to improve decision-making.

A Research of Fine-Tuning CNN-LSTM Model for Gold Price Prediction

Gold price and foreign exchange price prediction are usually regarded as serialized data with strong time correlation. It is precisely because of this feature that the performance of traditional convolutional neural networks is often not as good as that of recurrent neural networks. This paper fine-tunes the CNN-LSTM model in two ways. The first is to simplify the input data, convert the time series into binary data, and change the model task from predicting prices to predicting price fluctuations. The second is to add a dropout layer and use the dropout layer to improve the model's overfitting problem. The results demonstrate that the improved model outperforms the first model, with prediction accuracy rising from 54% to 57%. Furthermore, the refined model is not limited to processing the gold price series; it can also be applied to other time series, and when paired with other models, it could yield even better outcomes.

Road Crack Detection by Combining Dynamic Snake Convolution and Attention Mechanism

As one of the early manifestations of road pavement structure degradation, road cracks will accelerate the deterioration of the road if not detected and repaired in time. Aiming at the problems of low recall and incomplete crack detection in current road detection, based on the U-Net network, this paper proposed an Attention-Dynamic Snake Convolution U-Net (ADSC-U-Net) network. Firstly, the dynamic snake-shaped convolution was added to the normal downsampling process to make the network adaptively focus on the slender and curved local features, which can solve the problem of low accuracy of small crack detection. Secondly, the attention mechanism was used to pay better attention to the significant features of positive samples under the condition of a large proportion gap between positive and negative samples, which solved the problem of the poor crack integrity detection effect. Finally, the dataset was expanded by random vertical and horizontal flip operations, which solved the problem of network training overfitting caused by the small-scale datasets. The experimental results showed that, when the input image had a resolution of 480 × 320, evaluation indices P, R, and F1 of ADSC-U-Net on the self-built dataset were 74.44%, 68.77%, and 69.42%, respectively. Compared to SegNet, DeepLab, and DeepCrack, the P was improved by 1.90%, 2.49%, and 11.64%, respectively; the R was improved by 8.01%, 4.70%, and 59.58%, respectively; and the comprehensive evaluation index F1 was improved by 5.73%, 4.02%, and 55.87%, respectively, which proves the effectiveness of the proposed method.