Masked Regions Research Articles

SpheroScan: a user-friendly deep learning tool for spheroid image analysis.

In recent years, 3-dimensional (3D) spheroid models have become increasingly popular in scientific research as they provide a more physiologically relevant microenvironment that mimics in vivo conditions. The use of 3D spheroid assays has proven to be advantageous as it offers a better understanding of the cellular behavior, drug efficacy, and toxicity as compared to traditional 2-dimensional cell culture methods. However, the use of 3D spheroid assays is impeded by the absence of automated and user-friendly tools for spheroid image analysis, which adversely affects the reproducibility and throughput of these assays. To address these issues, we have developed a fully automated, web-based tool called SpheroScan, which uses the deep learning framework called Mask Regions with Convolutional Neural Networks (R-CNN) for image detection and segmentation. To develop a deep learning model that could be applied to spheroid images from a range of experimental conditions, we trained the model using spheroid images captured using IncuCyte Live-Cell Analysis System and a conventional microscope. Performance evaluation of the trained model using validation and test datasets shows promising results. SpheroScan allows for easy analysis of large numbers of images and provides interactive visualization features for a more in-depth understanding of the data. Our tool represents a significant advancement in the analysis of spheroid images and will facilitate the widespread adoption of 3D spheroid models in scientific research. The source code and a detailed tutorial for SpheroScan are available at https://github.com/FunctionalUrology/SpheroScan.

GSC-MIM: Global semantic integrated self-distilled complementary masked image model for remote sensing images scene classification

Masked image modeling (MIM) is a learning method in which the unmasked components of the input are utilized to learn and predict the masked signal, enabling learning from large amounts of unannotated data. However, due to the scale diversity and complexity of features in remote sensing images (RSIs), existing MIMs face two challenges in the RSI scene classification task: (1) If the critical local patches of small-scale objects are randomly masked out, the model will be unable to learn its representation. (2) The reconstruction of MIM relies on the visible local contextual information surrounding the masked regions and overemphasizing this local information will potentially lead the model to disregard the global semantic information of the input RSI. Regarding the above considerations, we proposed a global semantic integrated self-distilled complementary masked image model (GSC-MIM) for RSI scene classification. To prevent information loss, we proposed an information-preserved complementary masking strategy (IPC-Masking), which generates two complementary masked views for the same image to resolve the problem of masking critical areas of small-scale objects. To incorporate global information into the MIM pre-training process, we proposed the global semantic distillation strategy (GSD). Specifically, we introduced an auxiliary network pipeline to extract the global semantic information from the full input RSI and transfer the knowledge to the MIM by self-distillation. The proposed GSC-MIM is validated on three publicly available datasets of AID, NWPU-RESISC45, and UC-Merced Land Use, and the results show that the proposed method's Top-1 accuracy surpasses the baseline approaches in three datasets by up to 4.01, 3.87, and 5.26%, respectively.

Adaptive neuro-fuzzy inference system based data interpolation for particle image velocimetry in fluid flow applications

This paper presents an adaptive neuro-fuzzy inference system (ANFIS) approach for recovering the missing velocity vectors that commonly occur during fluid flow measurements in fluid mechanics. The capability of ANFIS in refilling the missing data is demonstrated with two case studies. First, the ANFIS is applied to estimate the velocity field data within the masked region of an available particle image velocimetry (PIV) experiment. Then, the ANFIS is trained using the data from outside the masked region and learns the relationship between the velocity vectors. Thus, it predicts the fluid patterns within the gappy area based on its understanding. ANFIS is also applied in another study to capture the small feature in the fluid flow. The vortices within a small area behind an obstacle are removed, and the rest of the data is introduced into the ANFIS model. The efficacy of the proposed ANFIS algorithm in predicting the missing velocity vectors is compared to both artificial neural network (ANN) and 2D cubic interpolation algorithms. It is found that intelligent algorithms (ANFIS and ANN) can predict the presence of vortices in the fluid flow even when there is no information about a circulating flow in the training dataset. The performance of ANFIS (R2= 0.91) in accurately predicting the velocity vectors is superior to the conventional ANN (R2= 0.86).

Large-area 2D selective area growth for photonic crystal surface emitting lasers

We report an investigation into large-area selective area growth of InGaAs/GaAs quantum wells by metalorganic vapour phase epitaxy. The emission wavelength tuning range, growth enhancement, and uniformity of material deposited within square masked regions with central square growth windows with widths in the range of 100–300 μm are studied. Micro-photoluminescence measurements at the centre point of each of the growth windows reveals a total wavelength tuning range of 86 nm across all samples, with a typical tuning range of 30 nm for a given window width, dependent upon dielectric mask width. The thickness enhancement in each of features, as determined by white-light interferometric profiling, indicates that centre point growth rate enhancements of between 1.19 and 2.23× are achieved with respect to the nominal epitaxial structure. By comparing the observed emission wavelengths with those simulated using the enhanced quantum well thicknesses, a range of indium concentrations between 12 and 17 % is calculated for the material at the centre of each feature. Two-dimensional analysis of selected features reveals that areas with uniform emission wavelength up to 100 × 100 μm2 in size can be achieved for the mask patterns used, indicating suitability for future applications in the fabrication of monolithically integrated multi-wavelength photonic crystal surface emitting laser arrays.

Citrus green fruit detection via improved feature network extraction.

It is crucial to accurately determine the green fruit stage of citrus and formulate detailed fruit conservation and flower thinning plans to increase the yield of citrus. However, the color of citrus green fruits is similar to the background, which results in poor segmentation accuracy. At present, when deep learning and other technologies are applied in agriculture for crop yield estimation and picking tasks, the accuracy of recognition reaches 88%, and the area enclosed by the PR curve and the coordinate axis reaches 0.95, which basically meets the application requirements.To solve these problems, this study proposes a citrus green fruit detection method that is based on improved Mask-RCNN (Mask-Region Convolutional Neural Network) feature network extraction. First, the backbone networks are able to integrate low, medium and high level features and then perform end-to-end classification. They have excellent feature extraction capability for image classification tasks. Deep and shallow feature fusion is used to fuse the ResNet(Residual network) in the Mask-RCNN network. This strategy involves assembling multiple identical backbones using composite connections between adjacent backbones to form a more powerful backbone. This is helpful for increasing the amount of feature information that is extracted at each stage in the backbone network. Second, in neural networks, the feature map contains the feature information of the image, and the number of channels is positively related to the number of feature maps. The more channels, the more convolutional layers are needed, and the more computation is required, so a combined connection block is introduced to reduce the number of channels and improve the model accuracy. To test the method, a visual image dataset of citrus green fruits is collected and established through multisource channels such as handheld camera shooting and cloud platform acquisition. The performance of the improved citrus green fruit detection technology is compared with those of other detection methods on our dataset. The results show that compared with Mask-RCNN model, the average detection accuracy of the improved Mask-RCNN model is 95.36%, increased by 1.42%, and the area surrounded by precision-recall curve and coordinate axis is 0.9673, increased by 0.3%. This research is meaningful for reducing the effect of the image background on the detection accuracy and can provide a constructive reference for the intelligent production of citrus.

SEM-RCNN: A Squeeze-and-Excitation-Based Mask Region Convolutional Neural Network for Multi-Class Environmental Microorganism Detection

This paper proposes a novel Squeeze-and-excitation-based Mask Region Convolutional Neural Network (SEM-RCNN) for Environmental Microorganisms (EM) detection tasks. Mask RCNN, one of the most applied object detection models, uses ResNet for feature extraction. However, ResNet cannot combine the features of different image channels. To further optimize the feature extraction ability of the network, SEM-RCNN is proposed to combine the different features extracted by SENet and ResNet. The addition of SENet can allocate weight information when extracting features and increase the proportion of useful information. SEM-RCNN achieves a mean average precision (mAP) of 0.511 on EMDS-6. We further apply SEM-RCNN for blood-cell detection tasks on an open source database (more than 17,000 microscopic images of blood cells) to verify the robustness and transferability of the proposed model. By comparing with other detectors based on deep learning, we demonstrate the superiority of SEM-RCNN in EM detection tasks. All experimental results show that the proposed SEM-RCNN exhibits excellent performances in EM detection.

Size Estimation of Mango Using Mask-RCNN Object Detection and Stereo Camera for Agricultural Robotics

—In agricultural robotics, to perform automatic harvesting task, it is important that the size of thetarget can be reasonably approximated.Stereo cameras can be used as a sensor to detect the targeted object and approximate its size. In this paper we present two methods for estimating the size of mangoes: pinhole model and ellipsoid model methods, using data obtained from a stereo camera.Object detection scheme using Mask-RCNN deep learning neural network on an RGB image and point cloud data from stereo camera are employed. The pinhole model method assumes simple triangular projection of the object onto the image plane.It makes use of the bounding boxes predicted by the neural network and the 3-D point cloud data. The ellipsoid model method, on the other hand, uses three-dimensional point cloud data obtained by processing the images from the stereo camera.It assumes that mango shape can be approximated as an ellipsoid and uses the predicted masked region and its corresponding point cloud data to fit an ellipsoid whose center and axes are extracted to estimate the position and the size of the mango, respectively. In the scope of this study where the full shape of mango can be seen in the image view, results from experiments suggest that the simpler pinhole model can effectively estimate the size of the mango with greater accuracy than the ellipsoid model. Keywords—Object Detection; Agricultural Robotics; Artificial Intelligent

Efficient Generative Face Completion with Perceptual Deblurring

Blurry textures in the inpainted regions or content inconsistencies between the inpainted and known regions may arise with existing deep-learning based completion approaches. A two-stage face completion framework with perceptual deblurring to address these problems is presented. In the first stage, a GAN is employed to inpaint masked regions with coarse facial features by integrating PSNR loss, SSIM loss and adversarial loss. In the second stage, a perceptual deblurring GAN is developed to further refine the coarsely reconstructed facial features in order to achieve clear facial textures and natural content consistencies. The quantitative, qualitative and ablation experiments are performed on the public face dataset CelebAHQ. Extensive experimental results show that the proposed algorithm is superior to the compared state-of-the-art image completion algorithms on the metrics of PSNR and SSIM.

Automatic Traffic Sign Recognition System Using CNN

In recent times, self-driving vehicles have been widely adopted across different countries as they are equipped to drastically reduce the number of road accidents and congestion on the road thereby improving the traffic efficiency. To detect, identify, and label the traffic signs on the road in order to help the Advanced Driver Assistance Systems (ADAS) in these autonomous vehicles with navigation details, a Traffic Sign Recognition (TSR) System using a deep convolutional neural network model, Mask RCNN (Mask Regional Convolutional Neural Network), is proposed in this paper that aims to help the autonomous vehicles comprehend the road ahead and safely navigate to the desired destination. This paper presents the detection and labelling of Indian and European Signs and also the results of the system working efficiently under various challenging visibility conditions. The results obtained show that the Mask RCNN model has recorded higher performance compared to all the other CNN models that have been previously used for traffic sign recognition.

In-field pose estimation of grape clusters with combined point cloud segmentation and geometric analysis

In-field object detection and pose estimation are challenging tasks in industrial harvesting scenarios. This study investigates a method for object detection and 6D pose estimation using a depth camera to prevent harvesting robots from collisions while harvesting grape clusters. First, the Mask Region Convolutional Neural Network (Mask R-CNN) is deployed to segment 2D images and output binary mask images of grape clusters. Second, the grape cluster point cloud is segmented based on the binary mask image and the mapping relationship between the image and point cloud to provide a high-quality point cloud through pre-processing. Third, the optimal cutting point is located by constructing the grape peduncle region of interest (RoI). Finally, the peduncle surface is fit based on the locally weighted scatterplot smoothing (LOWESS) algorithm, and the pose of the peduncle is estimated using geometric methods. The detection results from 238 test images show that the mean Precision (mP) is 86.0%, the mean Recall (mR) is 79.9%, the F1-score is 0.828, and mean Intersection Over Union (mIOU) for instance segmentation is 87.9%. The pose estimation results from 172 grape peduncles yield an error angle of 22.22 degrees ± 17.96 degrees in pose estimation. The detection and pose estimations for each grape cluster require approximately 1.786 s. The demonstrated performance of the proposed method indicates it can be applied to collision-free grape harvesting using robots in unstructured environments.

An approach for characterization of infected area in tomato leaf disease based on deep learning and object detection technique

Tomato leaf infections are a common threat to long-term tomato production that affects many farmers worldwide. Early detection, treatment, and solution of tomato leaf specificity are critical for promoting healthy tomato plant growth and ensuring ample supply and health security for the world’s geometric growth (population). The detection of plant leaf disease using computer-assisted technologies is prevalent these days. In this work, use the 1610 tomato leaf images of different classes from PlantVillage standard repository for the localization of objects. An effective Deep Learning (DL) modified Mask Region Convolutional Neural Network (Mask R-CNN) is proposed for the autonomous segmentation and detection of tomato plant leaf disease in this research. Intending to conserve memory space and computational expense, the suggested model adds a light head “Region Convolutional Neural Network (R-CNN)”. By varying the proportions of anchor in the RPN network and also changing the feature extraction topology, which improves the detection accuracy and computing the metric performance. The proposed technique is compared to existing state-of-the-art models to check if it is viable and robust. The outcomes of the suggested model achieved the results in terms of Mean Average Precision (mAP), F1-score, and accuracy of 0.88, 0.912, and 0.98, respectively. Furthermore, as the model’s ability increases with some parameters, the detection time for lesion detection is reduced by two times than the existing models.

"Deep Learning-based Classification of Materials into Biodegradable and Non-Biodegradable"

The collection of real-time data from people, their cars, public transit, buildings, and other urban infrastructures like the energy grid and waste management systems are at the heart of the smart city concept. The insights gathered from the data may be used by municipal authorities to manage resources and services efficiently. An important study topic at the same time is the sharp increase in environmental degradation and deterioration that causes ecological imbalance. Additionally, the development of advanced waste management systems that can categorize rubbish according to its level of biodegradability is required for the worldwide expansion of smart cities. Some of the more common ones are paper, paper boxes, food, glass, and other garbage. A costeffective technique to separate the waste from the enormous mountain of trash and garbage and classify the waste items is to use computer vision-based technologies. Recent developments in deep learning (DL) and deep reinforcement learning (DRL) have made it possible to categorize waste objects and identify and detect trash. In this sense, the research creates an intelligence model for smart cities. The technique's goal is to recognize and classify rubbish objects using the DL and DRL approaches. The two steps of the SYSTEM technique are objected classification based on DRL and object detection based on Mask Regional Convolutional Neural Network (CNN)… The CNN model uses the DenseNet model as a baseline model, and a deep learning network (DLN) is employed as the classifier. A based hyperparameter optimizer is also created to boost the efficiency of the DenseNet model.

Colposcopic multimodal fusion for the classification of cervical lesions

Objective: Cervical cancer is one of the two biggest killers of women and early detection of cervical precancerous lesions can effectively improve the survival rate of patients. Manual diagnosis by combining colposcopic images and clinical examination results is the main clinical diagnosis method at present. Developing an intelligent diagnosis algorithm based on artificial intelligence is an inevitable trend to solve the objectification of diagnosis and improve the quality and efficiency of diagnosis. Approach: A colposcopic multimodal fusion convolutional neural network (CMF-CNN) was proposed for the classification of cervical lesions. Mask region convolutional neural network was used to detect the cervical region while the encoding network EfficientNet-B3 was introduced to extract the multimodal image features from the acetic image and iodine image. Finally, Squeeze-and-Excitation, Atrous Spatial Pyramid Pooling, and convolution block were also adopted to encode and fuse the patient’s clinical text information. Main results: The experimental results showed that in 7106 cases of colposcopy, the accuracy, macro F1-score, macro-areas under the curve of the proposed model were 92.70%, 92.74%, 98.56%, respectively. They are superior to the mainstream unimodal image classification models. Significance: CMF-CNN proposed in this paper combines multimodal information, which has high performance in the classification of cervical lesions in colposcopy, so it can provide comprehensive diagnostic aid.

Comprehensive immune profiling unravels evolution of spatial distribution and immune repertoire in tumor microenvironment from primary to metastatic triple-negative breast cancer.

1079 Background: Immune checkpoint inhibitors (ICI) have improved PFS and OS in metastatic triple-negative breast cancer (mTNBC), but benefit is limited to PD-L1 positive tumors. Metastatic tumors are notorious for deficient immune cell infiltration and immunosuppressive features that may limit responses to ICI in mTNBC. However, the underlying mechanisms for the weak immunogenicity of the metastatic tumor immune microenvironment (TIME) and related poor ICI responses are still not well understood. The current study was designed to investigate the evolution of the TIME between paired primary and metastatic TNBCs. Methods: Spatial distribution of 37 key immune regulators using the NanoString digital spatial profiling (DSP) platform was analyzed. 452 regions of interests (ROIs) from 33 primary tumors (PT) and 29 metastatic tumors (MT) including 28 paired specimens, were selected based on CD45+ immune hotspots, and the protein expression levels of the key immune regulators were quantified within pan-cytokeratin (panCK) and CD45 masked regions, respectively. In parallel, we examined the clonality of tumor infiltrating B cell receptors by reconstructing the immune repertoire from bulk RNA-seq data. Results: Using the DSP platform, we confirmed reduced immune infiltration (e.g., CD3 and CD20) in both panCK and CD45 masked regions of MT, while CD8A and CD11c (padj = 2.8×10-7 and 2.1×10-6) expression was only observed in panCK masked regions of MT compared with PT. A significant shift in myeloid composition between PT and MT as evidenced by increased CD68 signal (padj = 5.8×10-4) in CD45 masked regions of MT was identified. Within MT, PD-L1 signal was substantially higher (padj = 0.030) in CD45 masked regions only, while PD1 counts were lower (padj = 0.035) in panCK masked regions. This suggests the limited responses to ICI for MT may stem from relatively low expression of activated and targetable T cell subsets in MT islands. In support of the lower CD20 counts in MT, immune repertoire analysis revealed B cell receptor (BCR) repertoire diversity (represented by Gini index) was substantially lower in MT than PT (p = 0.041) suggesting that the ability of B cells to recognize a wide variety of tumor antigens in MT is greatly reduced in contrast to PT. Conclusions: Through comprehensive analysis of the TIME spatial organization within paired PT and MT, a significant reduction in dendritic cell/macrophage ratios (CD11c/CD68), reduced tumor localized T cell activation (CD8, PD1, PD-L1), and reduced B cell diversity (BCR clonality) are key features of the reduced immunogenicity of the metastatic TIME in TNBC. Further work to understand key mechanistic features driving the evolution of these differences in TIME between primary and metastatic tumors are ongoing.

Object Detection via Gradient-Based Mask R-CNN Using Machine Learning Algorithms

Object detection has received a lot of research attention in recent years because of its close association with video analysis and image interpretation. Detecting objects in images and videos is a fundamental task and considered as one of the most difficult problems in computer vision. Many machine learning and deep learning models have been proposed in the past to solve this issue. In the current scenario, the detection algorithm must calculate from beginning to end in the shortest amount of time possible. This paper proposes a method called GradCAM-MLRCNN that combines Gradient-weighted Class Activation Mapping++ (Grad-CAM++) for localization and Mask Regional Convolution Neural Network (Mask R-CNN) for object detection along with machine learning algorithms. In our proposed method, images are used to train the network, together with masks that shows where the objects are in the image. A bounding box is regressed around the region of interest in most localization networks. Furthermore, just like any classification task, the multi-class log loss is minimized during training. This model enhances the calculation time and speed, as well as the efficiency, which recognizes objects in images accurately by comparing state-of-the-art machine learning algorithms, such as decision tree, Gaussian algorithm, k-means clustering, k-nearest neighbor, and logistic regression. Among these methods, we found logistic regression performed well with an accuracy rate of 98.4%, recall rate of 99.6%, and precision rate of 97.3% with respect to ResNet 152 and VGG 19. Furthermore, we proved the goodness of fit of our proposed model using chi-square statistical method and demonstrated that our solution can achieve great precision while maintaining a fair recall level.

Mask removal : Face inpainting via attributes

Due to the outbreak of the COVID-19 pandemic, wearing masks in public areas has become an effective way to slow the spread of disease. However, it also brings some challenges to applications in daily life as half of the face is occluded. Therefore, the idea of removing masks by face inpainting appeared. Face inpainting has achieved promising performance but always fails to guarantee high-fidelity. In this paper, we present a novel mask removal inpainting network based on face attributes known in advance including nose, chubby, makeup, gender, mouth, beard and young, aiming to ensure the repaired face image is closer to ground truth. To achieve this, a dual pipeline network based on GANs has been proposed, one of which is a reconstructive path used in training that utilizes missing regions in ground truth to get prior distribution, while the other is a generative path for predicting information in the masked region. To establish the process of mask removal, we build a synthetic facial occlusion that mimics the real mask. Experiments show that our method not only generates faces more similarly aligned with real attributes, but also ensures semantic and structural rationality compared with state-of-the-art methods.

Minkowski Functionals of SDSS-III BOSS: Hints of Possible Anisotropy in the Density Field?

Abstract We present measurements of the Minkowski functionals extracted from the SDSS-III BOSS catalogs. After defining the Minkowski functionals, we describe how an unbiased reconstruction of these statistics can be obtained from a field with masked regions and survey boundaries, validating our methodology with Gaussian random fields and mock galaxy snapshot data. From the BOSS galaxy data, we generate a set of four density fields in three dimensions corresponding to the northern and southern skies of LOWZ and CMASS catalogs, smoothing over large scales (Gaussian smoothing scale of 35 Mpc) such that the field is perturbatively non-Gaussian. We extract the Minkowski functionals from each data set separately, and measure their shapes and amplitudes by fitting a Hermite polynomial expansion. For the shape parameter of the Minkowski functional curves a 0, that is related to the bispectrum of the field, we find that the LOWZ-South data presents a systematically lower value of a 0 = −0.080 ± 0.040 than its northern sky counterpart a 0 = 0.032 ± 0.024. Although the significance of this discrepancy is not very high, it potentially indicates some systematics in the data or that the matter density field exhibits anisotropy at low redshift. By assuming a standard isotropic flat ΛCDM cosmology, the amplitudes of the Minkowski functionals from the combination of northern and southern sky data give the constraints Ωc h 2 n s = 0.110 ± 0.006 and 0.111 ± 0.008 for CMASS and LOWZ, respectively, which is in agreement with the Planck ΛCDM best-fit Ωc h 2 n s = 0.116 ± 0.001.

Restorable-inpainting: A novel deep learning approach for shoeprint restoration

Shoeprints are important information collected at the crime scene and are of great value for forensic analysis. Shoeprints collected in real-world scenarios are normally unclear, abrasive, and lack contextual and other kinds of missing information. In this research, we apply a novel deep learning technique called restorable inpainting to repair shoeprint contours and missing parts. Existing inpainting methods aim to fill artificially occluded areas with plausible pixels, but these methods may not restore missing information for occlusions in shoeprint images. In addition, because no ground-truth shoeprints exist for training samples, inpainting occluded regions becomes challenging. In this paper, we propose DeepShoePaint, a novel deep learning approach to perform restorable inpainting by restoring synthetic information resembling desirable shoeprint images necessary for forensics. DeepShoePaint novelly adapts a probabilistic distribution borrowed from the variational autoencoder into a U-Net-like structure forming a unified architecture trained in an unsupervised fashion to restore occluded and masked regions to produce human-verifiable shoeprints. The experimental results reveal that DeepShoePaint achieves outstanding human inspection and statistical assessment results and outperforms conventional inpainting models. We believe that this study can provide valuable insights, not limited to inpainting, into restoring desirable shoeprints to automate and facilitate the forensic investigation and examination process instead of using handcrafted methods.

Deep Learning-Based Computer-Aided Pneumothorax Detection Using Chest X-ray Images.

Pneumothorax is a thoracic disease leading to failure of the respiratory system, cardiac arrest, or in extreme cases, death. Chest X-ray (CXR) imaging is the primary diagnostic imaging technique for the diagnosis of pneumothorax. A computerized diagnosis system can detect pneumothorax in chest radiographic images, which provide substantial benefits in disease diagnosis. In the present work, a deep learning neural network model is proposed to detect the regions of pneumothoraces in the chest X-ray images. The model incorporates a Mask Regional Convolutional Neural Network (Mask RCNN) framework and transfer learning with ResNet101 as a backbone feature pyramid network (FPN). The proposed model was trained on a pneumothorax dataset prepared by the Society for Imaging Informatics in Medicine in association with American college of Radiology (SIIM-ACR). The present work compares the operation of the proposed MRCNN model based on ResNet101 as an FPN with the conventional model based on ResNet50 as an FPN. The proposed model had lower class loss, bounding box loss, and mask loss as compared to the conventional model based on ResNet50 as an FPN. Both models were simulated with a learning rate of 0.0004 and 0.0006 with 10 and 12 epochs, respectively.

Comparative Study on Concrete Crack Detection of Tunnel Based on Different Deep Learning Algorithms

The computer vision inspection of surface cracks in concrete structures has the characteristics of convenient and efficient 24-h on-site inspection. In this paper, the instance segmentation method and semantic segmentation method are used to realize the surface crack recognition of concrete structures, and a concrete crack detection method based on two different deep learning methods is designed. Experiments show that in scenarios where detection accuracy is required, compared with the Mask Region with Convolutional Neural Network method, the U-Net model improves the model segmentation accuracy through up-sampling and skip connection, and its Recall, Kappa, and Dice are increased by 6.88, 1.94, and 7.72%, respectively. In the scene of positioning requirements, Mask Region with Convolutional Neural Network has a better detection effect than U-Net for very thin and inconspicuous cracks, effectively avoiding the situation of missing cracks.