Human Inspection Research Articles

Defect classification of composite materials using transfer learning methods

ABSTRACT Nowadays, composite materials have become prevalent across various sectors, particularly finding usage in large-scale applications such as spaceships, automobiles, and aircrafts. The accurate detection of the defects in these materials is crucial, yet traditional methods often rely on human inspection, which is susceptible to errors. Recent advancements in machine learning have enabled defect detection using ultrasonic non-destructive testing methods. This paper introduces a new dataset named UNDT, which is obtained from the scans of 60 different composite materials, generating a total of 1150 images depicting both defective and non-defective areas. Several transfer learning methods are applied on the newly introduced UNDT dataset as well as the publicly available USimgAIST ultrasonic dataset. Comparative performance assessments illustrate the significance of utilising the transfer learning approach for defect classification on ultrasonic inspection images. Furthermore, the research emphasises the substantial benefits of employing these transfer learning methods. Notably, the DenseNet121 and VGG19 models achieve the highest accuracy rates, with 98.8% and 98.6% on the UNDT and USimgAIST datasets, respectively.

Cataclysmic variables from Sloan Digital Sky Survey - V (2020-2023) identified using machine learning.

Abstract SDSS-V is carrying out a dedicated survey for white dwarfs, single and in binaries, and we report the analysis of the spectroscopy of 504 cataclysmic variables (CVs) and CV candidates obtained during the first 34 months of observations of SDSS-V. We developed a convolutional neural network (CNN) to aid with the identification of CV candidates among the over 2 million SDSS-V spectra obtained with the BOSS spectrograph. The CNN reduced the number of spectra that required visual inspection to ≃ 2 per cent of the total. We identified 776 CV spectra among the CNN-selected candidates, plus an additional 27 CV spectra that the CNN misclassified, but that were found serendipitously by human inspection of the data. Analysing the SDSS-V spectroscopy and ancillary data of the 504 CVs in our sample, we report 61 new CVs, spectroscopically confirm 248 and refute 13 published CV candidates, and we report 82 new or improved orbital periods. We discuss the completeness and possible selection biases of the machine learning methodology, as well as the effectiveness of targeting CV candidates within SDSS-V. Finally, we re-assess the space density of CVs, and find 1.2 × 10−5 pc−3.

Enhanced Curvature-Based Fabric Defect Detection: A Experimental Study with Gabor Transform and Deep Learning

Quality control at every stage of production in the textile industry is essential for maintaining competitiveness in the global market. Manual fabric defect inspections are often characterized by low precision and high time costs, in contrast to intelligent anomaly detection systems implemented in the early stages of fabric production. To achieve successful automated fabric defect identification, significant challenges must be addressed, including accurate detection, classification, and decision-making processes. Traditionally, fabric defect classification has relied on inefficient and labor-intensive human visual inspection, particularly as the variety of fabric defects continues to increase. Despite the global chip crisis and its adverse effects on supply chains, electronic hardware costs for quality control systems have become more affordable. This presents a notable advantage, as vision systems can now be easily developed with the use of high-resolution, advanced cameras. In this study, we propose a discrete curvature algorithm, integrated with the Gabor transform, which demonstrates significant success in near real-time defect classification. The primary contribution of this work is the development of a modified curvature algorithm that achieves high classification performance without the need for training. This method is particularly efficient due to its low data storage requirements and minimal processing time, making it ideal for real-time applications. Furthermore, we implemented and evaluated several other methods from the literature, including Gabor and Convolutional Neural Networks (CNNs), within a unified coding framework. Each defect type was analyzed individually, with results indicating that the proposed algorithm exhibits comparable success and robust performance relative to deep learning-based approaches.

How Accurate are Transient Spectral Classification Tools?— A Study Using 4646 SEDMachine Spectra

Accurate classification of transients obtained from spectroscopic data are important to understand their nature and discover new classes of astronomical objects. For supernovae (SNe), SNID, NGSF (a Python version of SUPERFIT), and DASH are widely used in the community. Each tool provides its own metric to help determine classification, such as rlap of SNID, chi2/dof of NGSF, and Probability of DASH. However, we do not know how accurate these tools are, and they have not been tested with a large homogeneous data set. Thus, in this work, we study the accuracy of these spectral classification tools using 4646 SEDMachine spectra, which have accurate classifications obtained from the Zwicky Transient Facility Bright Transient Survey (BTS). Comparing our classifications with those from BTS, we have tested the classification accuracy in various ways. We find that NGSF has the best performance (overall Accuracy 87.6% when samples are split into SNe Ia and Non-Ia types), while SNID and DASH have similar performance with overall Accuracy of 79.3% and 76.2%, respectively. Specifically for SNe Ia, SNID can accurately classify them when rlap > 15 without contamination from other types, such as Ibc, II, SLSN, and other objects that are not SNe (Purity > 98%). For other types, determining their classification is often uncertain. We conclude that it is difficult to obtain an accurate classification from these tools alone. This results in additional human visual inspection effort being required in order to confirm the classification. To reduce this human visual inspection and to support the classification process for future large-scale surveys, this work provides supporting information, such as the accuracy of each tool as a function of its metric.

Intelligent Manufacturing in Wine Barrel Production: Deep Learning-Based Wood Stave Classification

Innovative wood inspection technology is crucial in various industries, especially for determining wood quality by counting rings in each stave, a key factor in wine barrel production. (1) Background: Traditionally, human inspectors visually evaluate staves, compensating for natural variations and characteristics like dirt and saw-induced aberrations. These variations pose significant challenges for automatic inspection systems. Several techniques using classical image processing and deep learning have been developed to detect tree-ring boundaries, but they often struggle with woods exhibiting heterogeneity and texture irregularities. (2) Methods: This study proposes a hybrid approach combining classical computer vision techniques for preprocessing with deep learning algorithms for classification, designed for continuous automated processing. To enhance performance and accuracy, we employ a data augmentation strategy using cropping techniques to address intra-class variability in individual staves. (3) Results: Our approach significantly improves accuracy and reliability in classifying wood with irregular textures and heterogeneity. The use of explainable AI and model calibration offers a deeper understanding of the model’s decision-making process, ensuring robustness and transparency, and setting confidence thresholds for outputs. (4) Conclusions: The proposed system enhances the performance of automatic wood inspection technologies, providing a robust solution for industries requiring precise wood quality assessment, particularly in wine barrel production.

Mechanics-informed autoencoder enables automated detection and localization of unforeseen structural damage

Structural health monitoring ensures the safety and longevity of structures like buildings and bridges. As the volume and scale of structures and the impact of their failure continue to grow, there is a dire need for SHM techniques that are scalable, inexpensive, can operate passively without human intervention, and are customized for each mechanical structure without the need for complex baseline models. We present a novel “deploy-and-forget” approach for automated detection and localization of damage in structures. It is a synergistic integration of entirely passive measurements from inexpensive sensors, data compression, and a mechanics-informed autoencoder. Once deployed, the model continuously learns and adapts a bespoke baseline model for each structure, learning from its undamaged state’s response characteristics. After learning from just 3 hours of data, it can autonomously detect and localize different types of unforeseen damage. Results from numerical simulations and experiments indicate that incorporating the mechanical characteristics into the autoencoder allows for up to a 35% improvement in the detection and localization of minor damage over a standard autoencoder. Our approach holds significant promise for reducing human intervention and inspection costs while enabling proactive and preventive maintenance strategies. This will extend the lifespan, reliability, and sustainability of civil infrastructures.

Advanced classification of hot subdwarf binaries using artificial intelligence techniques and Gaia DR3 data

Hot subdwarf stars are compact blue evolved objects, burning helium in their cores surrounded by a tiny hydrogen envelope. In the Hertzsprung-Russell Diagram they are located by the blue end of the Horizontal Branch. Most models agree on a quite probable common envelope binary evolution scenario in the Red Giant phase. However, the current binarity rate for these objects is yet unsolved, but key, question in this field. This study aims to develop a novel classification method for identifying hot subdwarf binaries within large datasets using Artificial Intelligence techniques and data from the third Gaia data release (GDR3). The results will be compared with those obtained previously using Virtual Observatory techniques on coincident samples. The methods used for hot subdwarf binary classification include supervised and unsupervised machine learning techniques. Specifically, we have used Support Vector Machines (SVM) to classify 3084 hot subdwarf stars based on their colour-magnitude properties. Among these, 2815 objects have Gaia DR3 BP/RP spectra, which were classified using Self-Organizing Maps (SOM) and Convolutional Neural Networks (CNN). In order to ensure spectral quality, previously to SOM and CNN classification, our 2815 BP/RP set were pre-analysed with two different approaches: the cosine similarity technique and the Uniform Manifold Approximation and Projection (UMAP) technique. Additional analysis onto a golden sample of 88 well-defined objects, is also presented. The findings demonstrate a high agreement level (sim 70-90<!PCT!>) with the classifications from the Virtual Observatory Sed Analyzer (VOSA) tool. This shows that the SVM, SOM, and CNN methods effectively classify sources with an accuracy comparable to human inspection or non-AI techniques. Notably, SVM in a radial basis function achieves 70.97<!PCT!> reproducibility for binary targets using photometry, and CNN reaches 84.94<!PCT!> for binary detection using spectroscopy. We also found that the single--binary differences are especially observable on the infrared flux in our Gaia DR3 BP/BR spectra, at wavelengths larger than $ sim $700 nm. We find that all the methods used are in fairly good agreement and are particularly effective to discern between single and binary systems. The agreement is also consistent with the results previously obtained with VOSA. In global terms, considering all quality metrics, CNN is the method that provides the best accuracy. The methods also appear effective for detecting peculiarities in the spectra. While promising, challenges in dealing with uncertain compositions highlight the need for caution, suggesting further research is needed to refine techniques and enhance automated classification reliability, particularly for large-scale surveys.

Implementation strategy for launch and performance improvement of high throughput manufacturing inspection systems

Product technologies are changing rapidly in advanced automotive propulsion systems. These products are driving the need for new manufacturing processes and new inspection methods. To keep new propulsion systems affordable and ensure these new products are introduced with high quality, automotive manufacturers are seeking automated inspection solutions with low cost and near-zero error rates to inspect 100% of the items. In this paper, a progressive deployment strategy of a hybrid inspection system is presented and studied in the context of technology development and rapid deployment. It enabled us to begin with human inspection and gradually phase-in automated inspection technology, while almost never failing to identify a bad item. This strategy was applied successfully to inspect ultrasonic welds in lithium ion battery packs. At the time of this study, a 75% reduction in human inspection was achieved with prospects for further reduction. Actual results from the implementation of this strategy in production are presented. Recommendations are made regarding the most appropriate time to employ this strategy and how it could increase the use of advanced automated in-line inspection technologies.

Achieving accurate prostate auto-segmentation on CT in the absence of MR imaging

BackgroundMagnetic resonance imaging (MRI) is considered the gold standard for prostate segmentation. Computed tomography (CT)-based segmentation is prone to observer bias, potentially overestimating the prostate volume by ∼ 30 % compared to MRI. However, MRI accessibility is challenging for patients with contraindications or in rural areas globally with limited clinical resources. PurposeThis study investigates the possibility of achieving MRI-level prostate auto-segmentation accuracy using CT-only input via a deep learning (DL) model trained with CT-MRI registered segmentation. Methods and MaterialsA cohort of 111 definitive prostate radiotherapy patients with both CT and MRI images was retrospectively grouped into training (n = 37) and validation (n = 20) (where reference contours were derived from CT-MRI registration), and testing (n = 54) sets. Two commercial DL models were benchmarked against the reference contours in the training and validation sets. A custom DL model was incrementally retrained using the training dataset, quantitatively evaluated on the validation dataset, and qualitatively assessed by two different physician groups on the validation and testing datasets. A contour quality assurance (QA) model, established from the proposed model on the validation dataset, was applied to the test group to identify potential errors, confirmed by human visual inspection. ResultsTwo commercial models exhibited large deviations in the prostate apex with CT-only input (median: 0.77/0.78 for Dice similarity coefficient (DSC), and 0.80 cm/0.83 cm for 95 % directed Hausdorff Distance (HD95), respectively). The proposed model demonstrated superior geometric similarity compared to commercial models, particularly in the apex region, with improvements of 0.05/0.17 cm and 0.06/0.25 cm in median DSC/HD95, respectively. Physician evaluation on MRI-CT registration data rated 69 %-78 % of the proposed model’s contours as clinically acceptable without modifications. Additionally, 73 % of cases flagged by the contour quality assurance (QA) model were confirmed via visual inspection. ConclusionsThe proposed incremental learning strategy based on CT-MRI registration information enhances prostate segmentation accuracy when MRI availability is limited clinically.

End-to-end solution for automatic beverage stock detection in supermarkets based on image processing and convolutional neural networks

This study addresses the challenge of detecting and identifying stock shortages in large warehouses through an advanced algorithm that integrates image processing and artificial intelligence techniques. Presently, many companies contend with the limitations of manual inventory management, such as susceptibility to errors, slow inventory actualizations, and consequent adverse economic effects. In contrast to solutions based on robotics, the proposed approach continuously monitors shelves throughout warehouse aisles using several fixed cameras, each connected to a single-board computer that processes the acquired images, identifies stock levels using deep learning, and updates a centralized database with stock analysis results. The algorithmic process begins with an image validation step based on a convolutional neural network to ensure obstacle-free images of the shelves. Subsequently, an application-specific YOLOv2 detector trained via transfer learning identifies product types captured in the images and estimates their stock levels. The proposed solution not only reduces the need for manual intervention and operational costs but also drastically enhances inventory supervision efficiency. The fully implemented system achieves an average accuracy of over 98 %, surpassing the human visual inspection performance. The proposed solution also incorporates the aspect of user-friendliness through a developed mobile application. This application connects to the centralized database, allowing inventory supervisors to receive alerts when the stock of a product falls below a user-configured threshold. This technological integration within a centralized system signifies a substantial advancement in inventory management, offering prompt responses to product scarcity situations and optimizing warehouse operational efficiency.

Automated particle inspection of continuously freeze-dried products using computer vision

The pharmaceutical industry is progressing towards more continuous manufacturing techniques. To dry biopharmaceuticals, continuous freeze drying has several advantages on manufacturing and process analytical control compared to batch freeze-drying, including better visual inspection potential. Visual inspection of every freeze-dried product is a key quality assessment after the lyophilization process to ensure that freeze-dried products are free from foreign particles and defects. This quality assessment is labor-intensive for operators who need to assess thousands of samples for an extensive amount of time leading to certain drawbacks. Applying Artificial Intelligence, specifically computer vision, on high-resolution images from every freeze-dried product can quantitatively and qualitatively outperform human visual inspection. For this study, continuously freeze-dried samples were prepared based on a real-world pharmaceutical product using manually induced particles of different sizes and subsequently imaged using a tailor-made setup to develop an image dataset (with particle sizes from 50μm to 1 mm) used to train multiple object detection models. You Only Look Once version 7 (YOLOv7) outperforms human inspection by a large margin, obtaining particle detection precision of up to 88.9% while controlling the recall at 81.2%, thus detecting most of the object present in the images, with an inference time of less than 1 s per vial.

Bird song comparison using deep learning trained from avian perceptual judgments.

Our understanding of bird song, a model system for animal communication and the neurobiology of learning, depends critically on making reliable, validated comparisons between the complex multidimensional syllables that are used in songs. However, most assessments of song similarity are based on human inspection of spectrograms, or computational methods developed from human intuitions. Using a novel automated operant conditioning system, we collected a large corpus of zebra finches' (Taeniopygia guttata) decisions about song syllable similarity. We use this dataset to compare and externally validate similarity algorithms in widely-used publicly available software (Raven, Sound Analysis Pro, Luscinia). Although these methods all perform better than chance, they do not closely emulate the avian assessments. We then introduce a novel deep learning method that can produce perceptual similarity judgements trained on such avian decisions. We find that this new method outperforms the established methods in accuracy and more closely approaches the avian assessments. Inconsistent (hence ambiguous) decisions are a common occurrence in animal behavioural data; we show that a modification of the deep learning training that accommodates these leads to the strongest performance. We argue this approach is the best way to validate methods to compare song similarity, that our dataset can be used to validate novel methods, and that the general approach can easily be extended to other species.

Critical Image Detection (CiD) and Single Image Detail Analysis (SiDA) – A practical approach to AI in NDT

Artificial Intelligence (AI) is gaining momentum in consumer and business applications and many approaches have changed the perspective on AI in Non-Destructive Testing (NDT). Most current discussions about AI in NDT revolve around automated defect detection / recognition (ADR), leaving the enormous potential of AI-based automation in other areas of the inspector's work process unexplored. It is in the nature of NDT that the human inspector's attention should be primarily focused on the parts that show indications or defects. These usually make up only a fraction of the components to be tested, as it is unknown which components actually show defects. This leads to a highly inefficient process that requires an evaluation of all components and thus ties up massive inspection capacities of human experts. The researchers of the German software company sentin have developed a system to process critical findings first and present those to the human inspector. NDT data such as images can be processed as batches leading to a (pre-) sorted dataset considering various indicators such as potential defects, duplicates, missing labels or references and many more. Such a resource-saving, high accuracy Critical Image Detection (CiD) system allows evaluators to find anomalies in complex and extensive datasets fast. After identification further automation can be achieved using a Single Image Detail Analysis system (SiDA), which is more resource intensive, but gives detailed information about the findings, potential defects and their location. The integration of these systems holds significant promises for the future application of AI in NDT. By streamlining the inspection process, they can generate substantial business value through time and cost savings.

Rapid counting of coliforms and Escherichia coli by deep learning‐based classifier

AbstractTo ensure that food has been handled hygienically, manufacturers routinely examine the numbers of indicator bacteria, such as coliforms and Escherichia coli. Using the deep‐learning algorithm YOLO, we developed a classifier that automatically counts the number of coliforms (red colonies) and E. coli (blue colonies) on a chromogenic agar plate. Using Citrobacter freundii IAM 12471T and E. coli NBRC 3301, we trained our YOLO‐based classifier with images of Petri dishes grown with each strain alone (10 images) and/or with a mixture of both strains (5 images). When the performance of the classifier was evaluated using 83 images, the accuracy rates for coliforms and E. coli reached 99.4% and 99.5%, respectively. We then investigated whether this classifier could detect other, non‐trained coliform species (22 species) and E. coli strains (13 strains). The accuracy rates for coliforms and E. coli were 98.7% (90 Petri dishes) and 94.1% (46 Petri dishes), respectively. Furthermore, we verified the practical feasibility of the developed classifier using 38 meats (chicken, pork, and beef). The accuracy rates for coliforms and E. coli in meat isolates were 98.8% (80 Petri dishes) and 93.8% (35 Petri dishes), respectively. The time required to count coliforms/E. coli on a single plate was ~70 ms. This novel method should enable users to rapidly quantify coliforms/E. coli without relying on a human inspector's color vision, leading to improved assurance of food safety.

Analyzing deep learning algorithms with statistical methods

A number of farmers use human inspection to classify the quality of their fruits visually; this process could be done automatically with a Learning algorithm. This paper investigates the classification of apples, peaches, and oranges and verifies the quality of classification with the density of the edges detected in each image by the Sobel filter. The novelty of the study consists of comparing the classification accuracy of a convolutional neural network (CNN) and the statistical analysis of edge histograms. The edges are the main elements of what the CNN learns; this was the reason for verifying the deep learning algorithms with skewness, kurtosis, and standard deviation metrics extracted from the histogram of edge density detected by the Sobel filter. In the empirical process, four algorithms were proposed: a CNN with one convolutional layer, a DNN without convolutional layers, and two pre-trained networks, AlexNet and GoogLeNet. From clustering of the row data and analyzing the accuracy, the best classification is obtained by simple CNN for 20 epochs when the apples vs. oranges were classified.

Usability Testing Via Simulation: Optimizing the NEAR4PEM Preintubation Checklist With a Human Factors Approach.

To inform development of a preintubation checklist for pediatric emergency departments via multicenter usability testing of a prototype checklist. This was a prospective, mixed methods study across 7 sites in the National Emergency Airway Registry for Pediatric Emergency Medicine (NEAR4PEM) collaborative. Pediatric emergency medicine attending physicians and senior fellows at each site were first oriented to a checklist prototype, including content previously identified using a modified Delphi approach. Each site used the checklist in 2 simulated cases: an "easy airway" and a "difficult airway" scenario. Facilitators recorded verbalization, completion, and timing of checklist items. After each simulation, participants completed an anonymous usability survey. Structured debriefings were used to gather additional feedback on checklist usability. Comments from the surveys and debriefing were qualitatively analyzed using a framework approach. Responses informed human factors-based optimization of the checklist. Fifty-five pediatric emergency medicine physicians/fellows (4-13 per site) participated. Participants found the prototype checklist to be helpful, easy to use, clear, and of appropriate length. During the simulations, 93% of checklist items were verbalized and more than 80% were completed. Median time to checklist completion was 6.2 minutes (interquartile range, 4.8-7.1) for the first scenario and 4.2 minutes (interquartile range, 2.7-5.8) for the second. Survey and debriefing data identified the following strengths: facilitating a shared mental model, cognitively offloading the team leader, and prompting contingency planning. Suggestions for checklist improvement included clarifying specific items, providing more detailed prompts, and allowing institution-specific customization. Integration of these data with human factors heuristic inspection resulted in a final checklist. Simulation-based, human factors usability testing of the National Emergency Airway Registry for Pediatric Emergency Medicine Preintubation Checklist allowed optimization prior to clinical implementation. Next steps involve integration into real-world settings utilizing rigorous implementation science strategies, with concurrent evaluation of the impact on patient outcomes and safety.

Detection of Unfocused EEG Epochs by the Application of Machine Learning Algorithm.

Electroencephalography (EEG) is a non-invasive method used to track human brain activity over time. The time-locked EEG to an external event is known as event-related potential (ERP). ERP can be a biomarker of human perception and other cognitive processes. The success of ERP research depends on the laboratory conditions and attentiveness of the test subjects. Specifically, the inability to control experimental variables has reduced ERP research in the real world. This study collected EEG data under various experimental circumstances within an auditory oddball paradigm experiment to enable the use of ERP as an active biomarker in normal laboratory conditions. Then, ERP epochs were analyzed to identify unfocused epochs, affected by typical artifacts and external distortion. For the initial comparison, the ability of four unsupervised machine learning algorithms (MLAs) was evaluated to identify unfocused epochs. Then, their accuracy was compared with the human inspection and a current EEG analysis tool (EEGLab). All four MLAs were typically 95-100% accurate. In summary, our analysis finds that humans might miss subtle differences in the regular ERP patterns, but MLAs could efficiently identify those. Thus, our analysis suggests that unsupervised MLAs perform better for detecting unfocused ERP epochs compared with the other two standard methods.

Neuroimaging article reexecution and reproduction assessment system.

The value of research articles is increasingly contingent on complex data analysis results which substantiate their claims. Compared to data production, data analysis more readily lends itself to a higher standard of transparency and repeated operator-independent execution. This higher standard can be approached via fully reexecutable research outputs, which contain the entire instruction set for automatic end-to-end generation of an entire article from the earliest feasible provenance point. In this study, we make use of a peer-reviewed neuroimaging article which provides complete but fragile reexecution instructions, as a starting point to draft a new reexecution system which is both robust and portable. We render this system modular as a core design aspect, so that reexecutable article code, data, and environment specifications could potentially be substituted or adapted. In conjunction with this system, which forms the demonstrative product of this study, we detail the core challenges with full article reexecution and specify a number of best practices which permitted us to mitigate them. We further show how the capabilities of our system can subsequently be used to provide reproducibility assessments, both via simple statistical metrics and by visually highlighting divergent elements for human inspection. We argue that fully reexecutable articles are thus a feasible best practice, which can greatly enhance the understanding of data analysis variability and the trust in results. Lastly, we comment at length on the outlook for reexecutable research outputs and encourage re-use and derivation of the system produced herein.

Two-Stage Detection Algorithm for Plum Leaf Disease and Severity Assessment Based on Deep Learning

Crop diseases significantly impact crop yields, and promoting specialized control of crop diseases is crucial for ensuring agricultural production stability. Disease identification primarily relies on human visual inspection, which is inefficient, inaccurate, and subjective. This study focused on the plum red spot (Polystigma rubrum), proposing a two-stage detection algorithm based on deep learning and assessing the severity of the disease through lesion coverage rate. The specific contributions are as follows: We utilized the object detection model YOLOv8 to strip leaves to eliminate the influence of complex backgrounds. We used an improved U-Net network to segment leaves and lesions. We combined Dice Loss with Focal Loss to address the poor training performance due to the pixel ratio imbalance between leaves and disease spots. For inconsistencies in the size and shape of leaves and lesions, we utilized ODConv and MSCA so that the model could focus on features at different scales. After verification, the accuracy rate of leaf recognition is 95.3%, and the mIoU, mPA, mPrecision, and mRecall of the leaf disease segmentation model are 90.93%, 95.21%, 95.17%, and 95.21%, respectively. This research provides an effective solution for the detection and severity assessment of plum leaf red spot disease under complex backgrounds.

Evaluating Human Expert Knowledge in Damage Assessment Using Eye Tracking: A Disaster Case Study

The rising frequency of natural disasters demands efficient and accurate structural damage assessments to ensure public safety and expedite recovery. Human error, inconsistent standards, and safety risks limit traditional visual inspections by engineers. Although UAVs and AI have advanced post-disaster assessments, they still lack the expert knowledge and decision-making judgment of human inspectors. This study explores how expertise shapes human–building interaction during disaster inspections by using eye tracking technology to capture the gaze patterns of expert and novice inspectors. A controlled, screen-based inspection method was employed to safely gather data, which was then used to train a machine learning model for saliency map prediction. The results highlight significant differences in visual attention between experts and novices, providing valuable insights for future inspection strategies and training novice inspectors. By integrating human expertise with automated systems, this research aims to improve the accuracy and reliability of post-disaster structural assessments, fostering more effective human–machine collaboration in disaster response efforts.