ABSTRACT Full‐field strain measurement is a critical technique for evaluating the mechanical properties and fracture behaviour of diverse materials. Imaging methodologies based on the sampling moiré (SM) method, which employ digital camera and grating patterns, offer remarkable versatility and effectiveness in strain analysis. In this study, we propose an improved SM method incorporating an image scaling technique for comprehensive strain measurement under varying deformation conditions. The proposed method was validated through simulations and tensile tests on angle‐ply carbon fibre reinforced polymer (CFRP) laminates, effectively capturing strain distributions during large deformation events, such as necking. Key results include precise strain distribution measurements beyond 2%, identification of necking deformation within the 2.5%–17% strain range and determination of material failure above 26.4% strain. These findings demonstrate the method's effectiveness and potential for advanced strain analysis, addressing limitations of conventional SM approaches in large deformation scenarios.

Software for drones has been created that allows tracking objects in real time. A software module has been developed that provides automatic capture and tracking of moving targets. The developed software uses computer vision algorithms and machine learning technologies. The object tracking mechanism has been integrated using CSRT or KCF trackers. An algorithm for automatic image scaling taking into account target movement has been implemented. Video stream processing has been optimized to ensure stable tracking of object movement in real time. Interactive control of the tracking process using control keys has been implemented.

Breast cancer detection and diagnosis remain challenging due to the complexity of tumor tissues and image quality variations, which hinder early and accurate identification. Timely diagnosis is vital for initiating treatment and improving patient outcomes. This study presents a novel hybrid feature fusion method, combining deep features from pre-trained models (ResNet50, InceptionV3, and MobileNetV2) with traditional texture features from Gabor filters and wavelet transforms, applied separately to mammogram and ultrasound datasets for breast cancer detection. A robust pre-processing pipeline, including image resizing, scaling, normalization, and CLAHE for contrast enhancement, is used to improve model performance. Data augmentation strengthens model robustness, and tumor segmentation is performed using Otsu’s multi-thresholding to accurately localize high-intensity regions. The hybrid feature extraction method yields 600 features, which are optimized through statistical feature selection for enhanced classification accuracy. Machine learning algorithms–XGBoost, AdaBoost, and CatBoost–are utilized to classify breast lesions across datasets, including Mini-DDSM and INbreast for mammogram images, and Rodrigues and BUSI for ultrasound images. Unlike most prior work, this fusion is applied across both mammogram and ultrasound modalities within one framework, a combination that has not been widely explored. Our approach explicitly targets the multi-modal gap to enhance robustness and generalizability across imaging types. The performance of the ensemble classifiers is compared, demonstrating the effectiveness of the proposed approach. The models achieved high classification accuracies: 98.67% for Rodrigues, 97.06% for INbreast, 97.02% for BUSI, and 95.00% for Mini-DDSM. These results highlight the effectiveness of the method and its potential to improve breast cancer detection. Future research will focus on comparisons with state-of-the-art models and real-world clinical applications.

Indonesia is one of the countries that is improving the quality of land data by digitizing. There are 14 million land parcels in Indonesia that have not been digitized (Ministry of ATR/BPN, 2024). Land data digitalization in Indonesia is still dominated by manual processes, especially in the digitization of land parcel boundaries on analog letter measurement. This results in high costs, lengthy processes, and the potential for human error in data interpretation. To solve this problem, deep learning technology can be utilized to automate the digitization of land parcel boundaries. One of them is by implementing the Segment Anything Model (SAM) algorithm. This research aims to develop a SAM-based User Interface (UI) capable of converting analog letter measurements of land parcel boundaries into digital spatial data. This UI is designed to facilitate users in processing and editing automation results. This research uses analog letter measurements from Mlati District, Sleman Regency, which contains 60 documents. This data is processed using the SAM algorithm to automate land parcel boundaries. The SAM algorithm has the role of automatically extracting land parcel boundary polygons with the assistance of simple prompts (points/boxes/text) from users. This process involves image enhancement, segmentation, regularization, and scaling to ensure that the automation results match the original documents. Evaluation was conducted by comparing the results of SAM automation with manual digitization using geometric testing methods. The results showed that the SAM algorithm can produce land parcel boundaries with very high accuracy, with precision, recall, F1 score, and accuracy values reaching 96% for simple geometry and 94% for complex geometry. Evaluation of the area still within the allowed tolerance of 5% (technical guidance PMNA/K.BPN No. 3 of 1997).

Abstract: Because plant diseases significantly affect crop output and food quality, they represent a major threat to global food security. Traditional manual methods for identifying diseases are labor-intensive, time-consuming, and often inaccurate. Convolutional Neural Networks (CNNs) are utilized in this AI-based plant health analyzer to detect and categorize plant leaf illnesses in order to get around these restrictions. The Plant Village dataset, which includes thousands of annotated photos of different crop species and disease types, is used to train the suggested system. For multi-class classification, the CNN model is implemented in PyTorch and optimized with the Adam optimizer with cross-entropy loss. Using a Flask-based web interface, users may enter leaf photographs and receive real-time diagnostic results, including the ailment diagnosis and confidence %. To improve model performance and generalization in a range of environmental conditions, a number of data augmentation techniques are employed, such as image flipping, rotation, and scaling. Experiments demonstrating exceptional disease prediction accuracy highlight the value of deep learning for precision agriculture. This clever technique facilitates early disease detection, allowing researchers and farmers to take prompt preventative action and supporting sustainable smart farming methods..

Radiographic measurements from spinal radiographs are crucial in many diagnostic and therapeutic decisions. However, widely used manual line drawing techniques in DICOM viewers are associated with significant errors, partly due to the unreliability of DICOM scale factors. A recently developed algorithm has been engineered to scale radiographs using assumed vertebral endplate widths (EPWs). Use of EPW to scale radiographs can eliminate the need to determine image magnification. This study was designed to (1) quantify the measurement error associated with the standard of care DICOM scale factor and (2) evaluate the efficacy of the EPW scaling algorithm. Previously collected cervical and lumbar lateral radiographs with a calibration marker of known size were used to collect radiographic measures of spinal alignment and spinal motion. Three different scale factors were used to acquire mm-based measurements: (1) the ground truth scale factor; (2) the DICOM scale factor; and (3) the EPW scale factor. DICOM and EPW scaled measurements were compared with respect to ground truth measurements. The DICOM scaled radiographic measurements demonstrated significantly more error compared to the EPW scaled radiographic measurements across multiple X-ray types: 19.6% vs. 9.7% in cervical neutral lateral (p < 0.001), 19.8% vs. 9.9% in cervical flexion/extension (p < 0.001), 35.0% vs. 7.5% in lumbar neutral lateral (p < 0.001), and 35.7% vs. 7.7% in lumbar flexion/extension (p < 0.001). The EPW algorithm provides adequate scaling accuracy for many clinical applications and is associated with considerably lower error than traditional DICOM scaled measurements. Clinicians should be aware of the potential scale factor inaccuracies in their imaging and be cautious when making diagnostic and therapeutic decisions based on improperly scaled radiographs.

Monkeypox has arisen as a global health issue, requiring prompt and precise diagnosis for optimal management. Conventional diagnostic techniques, including PCR, are dependable yet frequently unattainable in resource-constrained environments. Deep learning demonstrates potential in automating disease detection from skin lesion images; nevertheless, current models are hindered by limits in feature extraction and misclassification challenges. This paper presents an attention-augmented deep learning architecture to enhance classification accuracy for monkeypox and other dermatological conditions. This work presents a model based on EfficientNetB7, augmented with coordinate attention to enhance feature extraction and classification accuracy. The Monkeypox Skin Lesion Dataset (MSLD v2.0) is utilised, incorporating pre-processing methods such as image normalisation, scaling, and data augmentation. Diverse edge detection techniques are examined to enhance feature representation. The model is subjected to five-fold cross-validation and is evaluated against Xception, Swin Transformer, ResNet-50, MobileNetV2, and baseline EfficientNet models, utilising accuracy, precision, recall, F1-score, and AUC as assessment measures. Our model attains an unparalleled accuracy of 99.99%, precision of 99.8%, recall of 99.9%, F1-score of 99.85%, and an AUC of 100%. In contrast to previous studies that indicated a maximum accuracy of 98.81%, our methodology markedly diminishes false negatives and improves generalisation. This research sets a novel standard for AI-based monkeypox detection, showcasing exceptional accuracy and resilience. The results endorse the incorporation of AI-driven diagnostic tools in clinical and telemedicine settings, with prospects for immediate implementation and extensive epidemiological monitoring.

Background: Segmenting nerve fibres in histological images is a tricky job because of how much the tissue looks can change. Modern neural network architectures, including U-Net and transformers, demonstrate varying degrees of effectiveness in this area. The aim of this study is to conduct a comparative analysis of the SegFormer, VGG-UNet, and FabE-Net models in terms of segmentation quality and speed. Methods: The training sample consisted of more than 75,000 pairs of images of different tissues (original slice and corresponding mask), scaled from 1024 × 1024 to 224 × 224 pixels to optimise computations. Three neural network architectures were used: the classic VGG-UNet, FabE-Net with attention and global context perception blocks, and the SegFormer transformer model. For an objective assessment of the quality of the models, expert validation was carried out with the participation of four independent pathologists, who evaluated the quality of segmentation according to specified criteria. Quality metrics (precision, recall, F1-score, accuracy) were calculated as averages based on the assessments of all experts, which made it possible to take into account variability in interpretation and increase the reliability of the results. Results: SegFormer achieved stable stabilisation of the loss function faster than the other models-by the 20-30th epoch, compared to 45-60 epochs for VGG-UNet and FabE-Net. Despite taking longer to train per epoch, SegFormer produced the best segmentation quality, with the following metrics: precision 0.84, recall 0.99, F1-score 0.91 and accuracy 0.89. It also annotated a complete histological section in the fastest time. Visual analysis revealed that, compared to other models, which tended to produce incomplete or excessive segmentation, SegFormer more accurately and completely highlights nerve structures. Conclusions: Using attention mechanisms in SegFormer compensates for morphological variability in tissues, resulting in faster and higher-quality segmentation. Image scaling does not impair training quality while significantly accelerating computational processes. These results confirm the potential of SegFormer for practical use in digital pathology, while also highlighting the need for high-precision, immunohistochemistry-informed labelling to improve segmentation accuracy.

ABSTRACT The traditional methods of monitoring and detecting diseases in crops are being replaced by automated techniques that have the potential to revolutionize the innovative agriculture industry. Guava, a crop of paramount importance in Asian countries, is unfortunately susceptible to disease and fungal attacks that significantly impact production. This study utilized renowned Convolutional Neural Networks (CNN) to diagnose diseases on guava fruit and leaves. The data set used in the study, derived from Mendeley data, encompassed four types of guava diseases and healthy ones. The data underwent preprocessing, including image scaling, data augmentation techniques, and a Threshold-Based pixel filtering approach to focus on Guava. The study's findings revealed that EfficientNetB3 outperformed AlexNet, ResNet50, various versions of MobileNet, and previous studies in Guava Disease Identification tasks.

Sign language recognition is an essential tool that facilitates communication for those with hearing and speech disabilities. Conventional recognition techniques frequently encounter challenges in real-time performance, resilience, and accuracy owing to fluctuations in hand positions, backdrops, and lighting conditions. This paper presents a YOLOv11-based deep learning system for recognising American Sign Language (ASL), concentrating on both alphabetic and transactional hand motions to mitigate existing constraints. The model is engineered to function in real-time while ensuring high precision and resilience across varied contexts. The methodology adheres to a systematic pipeline, commencing with dataset gathering and pre-processing, which include image augmentation, normalisation, and scaling to guarantee model generalisation. The YOLOv11 architecture utilises an improved backbone, neck, and detecting head for effective feature extraction and classification. Training is enhanced by the utilisation of the AdamW optimiser, a meticulously adjusted learning rate, and a loss function that integrates box loss, classification loss, and distribution focal loss (DFL). Performance is assessed using precision, recall, mean Average Precision (mAP), and inference rate to guarantee the model's accuracy and efficiency. Experimental findings indicate that the suggested model attains 95.4% precision, 94.8% recall, and 98.1% mean Average Precision (mAP), markedly surpassing conventional methods. The amalgamation of GRAD-CAM with occlusion sensitivity significantly improves model interpretability. This research offers a robust and scalable approach for real-time sign language detection, facilitating enhanced accessibility in communication technologies, assistive devices, and interactive systems.

The paper presents a mathematical model for real-time object detection and recognition in video streams, based on stepwise analysis of inter-frame changes. The proposed approach integrates basic linear and morphological operations with an efficient inter-frame differencing procedure, enabling the localization of moving or newly appearing objects across consecutive frames, followed by their classification using neural networks. The formalized algorithmic structure of the model covers all essential stages: image scaling, grayscale conversion, absolute difference computation, threshold filtering, morphological cleanup, extraction of regions of interest, object classification, and subsequent temporal tracking. The model is structured as a sequence of functional transformations addressing both spatial and temporal aspects of video data processing. The use of inter-frame differencing as a core activity detector is justified as it significantly reduces the computational burden in comparison with fully convolutional deep learning models such as SSD or YOLO. Classical morphological filters (opening and closing) are employed to refine object contours, while size-based region filtering helps exclude noisy or irrelevant areas. At the final stage, validated regions are passed to a classification module, allowing identification of object types and enabling tracking without repeated detection. An experimental evaluation was conducted using footage from a static camera to assess the model’s effectiveness. The results demonstrate an average frame processing time of 5.4 ms, meeting real-time operational requirements, and a recognition accuracy of 71.2%. Profiling indicates that the most computationally intensive operations are associated with morphological processing, whereas classification accounts for less than half of the total processing time. This highlights the efficiency of the hybrid approach, where simple linear preprocessing significantly reduces the data load for classification without substantial accuracy loss.

DNA-based storage has emerged as a promising storage paradigm due to its immense storage potential. However, the error-prone nature of DNA sequencing and synthesis processes limits this potential. Image data is typically compressed before storage, and even a single mismatch can lead to catastrophic error propagation during decompression, rendering the image unrecoverable. To reduce the error rate of DNA storage-based image compression, we have designed a high fault-tolerant DNA image storage system and applied it to image compression for DNA storage. This system achieves significant improvements in both image data compression ratio and resilience through three key innovations: 1) Using a Variational Autoencoder (VAE) to compress the image into uniformly sized latent variable blocks, followed by further compression via Singular Value Decomposition (SVD); 2) Quantizing the floating-point numbers in the latent variable blocks and applying rotational coding to the resulting ternary sequences, effectively ensuring positive constraints on homopolymer run lengths and GC content; 3) Optimizing the error-correction scheme to best recover each type of error by quantizing it back to its original value. Through image scaling, we adjust the compression ratio, and the comparative results of image compression simulations demonstrate the performance of the proposed model, highlighting its superiority in fault tolerance and storage density.

The hardware and software of a telecommunication system designed for the transmission of adaptively scaled digital images has been developed. The system consists of transmission and reception subsystems. The hardware of the transmission subsystem includes a USB video camera, a microcomputer Raspberry Pi3 and a Raspberry Pi3 radio module. The hardware of the reception subsystem includes a radio module nRF24L01 # 2, a microcontroller Funduino Uno and a computer. The radio module Funduino Uno has a relatively low transmission speed, but provides low power consumption. The system software has been developed in Python and C++. In the transmission subsystem, the image from the video camera is read using the program cam2nrf, its scale is reduced by interpolation and the reduced image is transmitted via radio module # 1. In the reception subsystem, the reduced image is read via radio module # 2, its scale is increased by interpolation. Image scaling is performed using bilinear or bicubic interpolation. The program nrf_rx, which runs on the Funduino Uno microcontroller, is used to receive wireless data using the radio module nRF24L01. The program recv_img, which runs on the computer, reads the image from the microcontroller Funduino Uno, scales, visualizes and saves it. Scaling adaptability is ensured by choosing the image interpolation algorithm depending on its average spatial period TCR, which is calculated based on the Fourier power spectrum of the original image. For values TCR ​​&lt; 4.5 pixels, bilinear interpolation is performed, and otherwise bicubic interpolation is performed. Testing of the telecommunication system for transmitting real images shows its operability. By reducing the image size by a factor of 2, the transmission time is reduced by a factor of 4 with a slight decrease in the visual quality of the resulting image. Such image scaling is especially effective when transmitting images over low-bandwidth channels. The developed hardware and software can be used in `Internet of Things systems for image transmission.

Image scaler is widely used to convert an image from one spatial resolution to another resolution. For the display driver integrated circuits (DDIC) of a display panel, image scaler is needed to serve as an individual IP. Since image data is sent to display buffer in a row‐by‐row way, just a few (usually no more than 4) rows of image data can be used for image scaling, rather than the whole image data, to save memory size and cost for DDIC. In this work, we propose a high‐quality, low‐cost image scaler with less line buffer for display in mobile devices. A four‐line buffer is used to decide the periodicity of image data. Then, non‐periodic and periodic samples are separately treated. For periodic samples, they are directly scaled via periodic sample register bank architecture. For non‐periodic samples, namely texture image data, an adaptive edge‐enhanced image scaling is used with an improved image sharpening filter with 3x3 kernel. The proposed image scaler can be implemented by a simple hardware architecture with fewer multipliers and adders, consuming less hardware area. Moreover, it achieves much better scaling for images with periodic patterns.

Acrophobia, the intense fear of heights, affects a significant portion of the global population. Traditional therapeutic approaches, such as Cognitive Behavioral Therapy (CBT), have been supplemented by emerging technological solutions like Virtual Reality (VR). This paper explores the role of VR-based visual imagery in treating acrophobia. We examine the use of immersive environments, integration of multi-sensory stimuli, and how images in VR are designed to elicit controlled exposure therapy outcomes. Additionally, the paper discusses the impact of realism, scaling, and interactivity of VR-generated images on patient treatment efficacy. VR-based acrophobia treatment, particularly when leveraging realistic images and immersive environments, represents a promising advancement in therapeutic techniques for phobia management.

Film-specific uniformity variations in packages are known to significantly diminish the effectiveness of the one-scan protocol, a commonly used film dosimetry method. This method universally adopts the reference dose-response with rescaling linearly from the relationship of the known dose and the unexposed state. This study aims to visualize and quantify the variation in unexposed film-specific uniformity in a package to evaluate the suitability of the reference dose response using machine-learning method. Fourteen EBT4 films (#00-#13) were selected from two lot packages. Nine grid-spaced 100 × 100 pixel (72 dpi) patches were obtained from the color images of EBT4 film sheet using a single scanner with landscape (scan A) and portrait (scan B) scan orientations. The reference patch was set at the center of film #00. For this study, multidimensional scaling (MDS) and Lie derivative image analysis (LDIA) were applied to the patch data with respect to the red (R)/green (G)/blue (B) channels. MDS is a suitable method for analyzing non-linear data with similarity, which provides a map of data objects according to a distance metric. LDIA directly detects the deviation vector field between image gradients. The film-specific uniformity was measured at 1/10000 scaled pixel value as a scalar distribution. The image flow field was obtained as a negative gradient of the scalar distribution. Two similarity metrics were defined for comparison with the reference patch: (1) MDSr (the distance parameter in the MDS map from the origin) and (2) Stot (summed S-value in each patch, where S-value represents the vorticity of the deviation vector field obtained via the Lie derivative). MDSr highly correlated with the absolute pixel value difference from the reference patch except for the blue channel in which a favorable package was detected for the reference dose response. Stot quantified the film-uniformity variation from the reference, independent of the dataset, and detected the unfavorable film state as Stot < 0.8 in the blue channel. We visualized and quantified the variation in film-specific uniformity in a lot package using MDS and LDIA, thereby quantitatively determining the unfavorable condition for applying the reference dose-response.

This project aims to develop a weed detection prototype for agricultural settings using the Convolutional Neural Networks (CNN) algorithm. The project thoroughly analyses and optimises CNN hyperparameters to improve accuracy and efficiency, empowering efficient weed control practices. The potential of this algorithm in weed detection is immense, offering a promising future for sustainable productivity in agriculture. Adopting innovative and sustainable agricultural practices is essential for building a robust and productive agriculture sector that can meet future food demands while protecting the environment. The research then assesses how well the CNN model generalises to various agricultural environments that support multiple crop situations. The dataset comprises 360 images of weeds, broadleaf, maise plants, soil and cotton crops. The images underwent four preprocessing phases: image scaling, normalisation, filtering, and segmentation. The proposed model achieved an accuracy of 89.82% utilizing the Convolutional Neural Network (CNN) algorithm, with the dataset partitioned into 80% for training and 20% for testing. Furthermore, the model attained an F1 score of 88.08%, indicating a high degree of alignment between predicted positive instances and actual positive samples. In addition to technological innovations in agriculture, this CNN-based weed detection prototype is a reliable resource for agriculturalists. AI-driven weed detection optimizes resource use, ensuring that pesticides and herbicides are applied only where necessary, reducing chemical overuse. This is in line with the United Nation Sustainable Development Goal (SDG) No. 12.

Enhanced function approximation and applications to image scaling: A new family of exponential sampling neural network Kantorovich operators

The previous advancements in pathology image understanding primarily involved developing models tailored to specific tasks. Recent studies have demonstrated that the large vision-language model can enhance the performance of various downstream tasks in medical image understanding. In this study, we developed a domain-specific large vision-language model (PathologyVLM) for pathology image understanding. Specifically, (1) we first construct a human pathology image-text dataset by cleaning the public medical image-text data for domain-specific alignment; (2) Using the proposed image-text data, we first train a pathology language-image pretraining (PLIP) model as the specialized visual encoder to extract the features of pathology image, and then we developed scale-invariant connector to avoid the information loss caused by image scaling; (3) We adopt two-stage learning to train PathologyVLM, first stage for domain alignment, and second stage for end to end visual question & answering (VQA) task. In experiments, we evaluate our PathologyVLM on both supervised and zero-shot VQA datasets, our model achieved the best overall performance among multimodal models of similar scale. The ablation experiments also confirmed the effectiveness of our design. We posit that our PathologyVLM model and the datasets presented in this work can promote research in field of computational pathology. All codes are available at: https://github.com/ddw2AIGROUP2CQUPT/PA-LLaVA

Deep learning-based target detection algorithms outperform traditional methods by eliminating the need for manual feature design and improving accuracy and efficiency. This paper constructs a YOLOv5 target detection model using a deep convolutional neural network. To enhance accuracy, generalization, and detection speed, three data augmentation techniques—mosaic data enhancement, adaptive anchor frame, and adaptive image scaling—are applied. The model is further optimized with an attention mechanism and a modified YOLOv5 framework. A loss function and global average pooling enhance feature mapping for a fully convolutional network. Experimental results show that the improved YOLOv5n model achieves a 2.9979 percentage point increase in MAP, a 31% improvement in FPS, and a training time reduction of 10 minutes, completing 100 rounds in 20 minutes.