ResNet Architecture Research Articles

A Comparative Study of Fine-Tuning Deep Learning Models for Leaf Disease Identification and Classification

Innovative agricultural solutions are needed to detect and classify leaf diseases early across crop species and environments. This study compares deep learning approaches, focusing on Convolutional Neural Networks (CNN) and Vision Transformers (VTs), to identify leaf diseases early and accurately for scalable crop management and productivity. Optimizing CNNs, Explainable Transfer Learning (ETPLDNet) using ResNet50 architecture, and LEViT leaf disease diagnosis are compared. The CNN model, optimized with dynamic hyperparameters, achieved an impressive 99.58% accuracy for leaf disease classification, demonstrating its effectiveness in feature extraction and classification precision. On the other hand, the VT-based LEViT model, which leverages self-attention mechanisms and Explainable AI (XAI), achieved 95.22% accuracy but offers enhanced interpretability and generalization capabilities due to its transformer-based architecture. This distinction illustrates that while CNNs excel in accuracy, VTs provide a more transparent decision-making process and better handle the complex variances in plant leaf diseases, making them ideal for precision agriculture. The combined use of CNNs and VTs showcases the strengths of each model, with CNN focusing on high classification precision and VTs offering improved interpretability and adaptability for various leaf disease conditions. The use of XAI enables the models to highlight important areas in plant leaf images that influence the model's decisions, offering a transparent and interpretable decision-making process that allows researchers and farmers to understand why a particular diagnosis or classification was made. This ability to visualize and explain the reasoning behind the model predictions is crucial to increasing trust in AI-driven solutions in agriculture. By combining the high precision of CNN and the interpretability of VT with XAI, this study offers a robust approach to improving crop disease management and precision agriculture.

Convolutional neural network-based wire-cut image recognition and defect detection research

Abstract This paper presents an advanced defect detection system for wire Electrical Discharge Machining (EDM), utilizing Convolutional Neural Networks (CNNs) to automatically identify, classify, and localize defects such as cracks, notches, and burrs. Wire EDM is a precision manufacturing process critical for cutting conductive materials, where defect detection plays a vital role in ensuring product quality. The proposed method incorporates a modified ResNet-50 architecture, optimized specifically for defect detection in wire EDM. The architecture leverages deep residual learning to enhance feature extraction, allowing the system to detect minute defects effectively. A dataset of 10,000 RGB images (224x224 pixels) was used for training, with the model achieving an impressive 95.3% accuracy, 94.2% recall, 95.8% precision, and 94.7% F1 score on the test set. The system demonstrated excellent performance in detecting cracks, though it showed slightly lower performance on deformation-related defects. A comprehensive comparison with traditional defect detection methods and other deep learning models underscores the superiority of the proposed approach in terms of both accuracy and robustness. The results indicate that this CNN-based system offers a reliable and efficient solution for quality control in wire EDM processes. Future research will focus on further optimizing the network for real-time defect detection and extending the approach to incorporate multi-modal data, such as sensor and acoustic signals, to improve overall detection performance.

MammoViT: A Custom Vision Transformer Architecture for Accurate BIRADS Classification in Mammogram Analysis

Background: Breast cancer screening through mammography interpretation is crucial for early detection and improved patient outcomes. However, the manual classification of mammograms using the BIRADS (Breast Imaging-Reporting and Data System) remains challenging due to subtle imaging features, inter-reader variability, and increasing radiologist workload. Traditional computer-aided detection systems often struggle with complex feature extraction and contextual understanding of mammographic abnormalities. To address these limitations, this study proposes MammoViT, a novel hybrid deep learning framework that leverages both ResNet50’s hierarchical feature extraction capabilities and Vision Transformer’s ability to capture long-range dependencies in images. Methods: We implemented a multi-stage approach utilizing a pre-trained ResNet50 model for initial feature extraction from mammogram images. To address the significant class imbalance in our four-class BIRADS dataset, we applied SMOTE (Synthetic Minority Over-sampling Technique) to generate synthetic samples for minority classes. The extracted feature arrays were transformed into non-overlapping patches with positional encodings for Vision Transformer processing. The Vision Transformer employs multi-head self-attention mechanisms to capture both local and global relationships between image patches, with each attention head learning different aspects of spatial dependencies. The model was optimized using Keras Tuner and trained using 5-fold cross-validation with early stopping to prevent overfitting. Results: MammoViT achieved 97.4% accuracy in classifying mammogram images across different BIRADS categories. The model’s effectiveness was validated through comprehensive evaluation metrics, including a classification report, confusion matrix, probability distribution, and comparison with existing studies. Conclusions: MammoViT effectively combines ResNet50 and Vision Transformer architectures while addressing the challenge of imbalanced medical imaging datasets. The high accuracy and robust performance demonstrate its potential as a reliable tool for supporting clinical decision-making in breast cancer screening.

A Deep Learning-Based Approach to Detect Lamina Dura Loss on Periapical Radiographs.

This study aimed to develop a custom artificial intelligence (AI) model for detecting lamina dura (LD) loss around the roots of anterior and posterior teeth on intraoral periapical radiographs. A total of 701 periapical radiographs of the anterior and posterior regions retrieved from the Dentomaxillofacial Radiology archives were reviewed. Images were cropped to include only the teeth exhibiting LD loss and those without LD loss, which were labeled as "1" and "0," respectively. The dataset was diversified using image preprocessing and data augmentation techniques. Among the radiographs, 72% were used for training, 18% for validation, and 10% for testing. A custom AI model, consisting of 4 blocks and 49 layers, with a total of 21.2 million parameters, was developed using the TensorFlow library and residual blocks introduced in ResNet architecture. Sensitivity, specificity, accuracy, precision, F1 score, and kappa (κ) coefficients (for intra-observer agreement) were calculated to evaluate the performance of the AI model. When applied to a test set of 71 images, the AI model showed good performance in detecting LD loss, achieving an average sensitivity of 0.730, specificity of 0.706, accuracy of 0.718, precision of 0.730, and an F1 score of 0.730, regardless of the dental region. This study represents the first known application of an AI algorithm tailored to detect LD loss on periapical radiographs. The developed AI model could aid clinicians in making accurate diagnosis and help prevent misdiagnosis.

EEG-Based ADHD Classification Using Autoencoder Feature Extraction and ResNet with Double Augmented Attention Mechanism.

Attention-Deficit/Hyperactivity Disorder (ADHD) represents a widely prevalent and heterogeneous neurodevelopmental condition in pediatric populations, often exhibiting a substantial propensity to persist into adulthood. ADHD is a multifaceted disorder that resists straightforward diagnostic tests. Clinicians must invest substantial time and effort to secure an accurate diagnosis and implement effective treatment. ADHD diagnosis is primarily based on psychiatric tests, as there is currently no clinically utilized objective diagnostic tool. Nonetheless, several studies in have documented endeavors to create objective instruments designed to assist in the diagnostic process of ADHD, aiming to enhance diagnostic accuracy and reduce subjectivity. This research endeavor sought to establish an objective diagnostic modality for ADHD through the utilization of electroencephalography (EEG) signal analysis. With the use of innovative deep learning techniques, this research seeks to improve the diagnosis of ADHD using EEG data. To capture complex patterns in EEG data, this study proposes a double-augmented attention mechanism ResNet-based model. Using an autoencoder for feature extraction, the Reptile Search Algorithm for feature selection, and a modified ResNet architecture for model training comprise the technique. AUC, F1-score, accuracy, precision, recall, and other standard classifiers like Random Forest and AdaBoost were utilized to compare the model's performance. By a wide margin, the proposed ResNet model outperforms the traditional models with a 99.42% accuracy, 99.03% precision, 99.82% recall, and 99.42% F1-score. ROC AUC score of 0.99 for the model underscores its remarkable capability to differentiate between children with and without ADHD, thereby minimizing misclassification errors and improving diagnostic precision.

Intelligent classification of lung cancer pathology images through comparative morphological feature learning

Lung cancer pathology images are categorized with enhanced accuracy in this research, addressing the significant challenge posed by the limited availability of labeled images, a limitation exacerbated by the complexity of cellular morphologies. Background The accurate classification of lung cancer pathology images is of paramount importance for both diagnostic and therapeutic purposes. However, the development of robust classification models is often hindered by the intricate cellular morphologies and the scarcity of labeled images, which is a critical bottleneck in the field. Objectives The study is designed to incorporate unlabeled data into the training process, thereby enhancing the classification of lung cancer pathology images through the use of comparative learning techniques. Methods A methodology is introduced wherein confidently classified unlabeled images are integrated with labeled ones, enriching the training dataset. This approach draws on principles of farthest and nearest neighbor contrastive learning to cultivate a more challenging learning environment and to augment the variability of contrastive samples. To effectively extract key cellular morphological features, an encoder based on the ResNet50 architecture, fortified with deformable and dynamic convolutional techniques, is utilized. Results Demonstrated by experimental results, the proposed classification strategy achieves a significant improvement in the accuracy of lung cancer image classification, even under conditions characterized by a limited availability of labeled data, thus underscoring the robustness of the method. Conclusion The integration of comparative learning with both labeled and unlabeled images, complemented by the application of advanced convolutional techniques, is shown to be a promising avenue for enhancing the classification of lung cancer pathology images. This research is presented as a practical solution to the urgent need for accurate and efficient diagnostic tools in the field of oncology.

Thermal Image Classification of Autistic Children Using Res-Net Architecture

The thermal Image Classification Method has been widely used for significant applications in many fields, including thermal images of the face. This study presents a method for thermal facial classification in children with autism spectrum disorder (ASD). Children with ASD have a neurological disorder that affects communication skills essential in daily life and often causes difficulties in social situations. As we know, the diagnosis of ASD currently still relies on human methods and does not yet have definite biological markers. Early diagnosis of ASD has a significant positive impact, especially in children. Deep learning techniques, especially in facial medical image analysis, have become a new research focus in ASD detection. Initial screening using a Convolutional Neural Network (CNN) model with a transfer learning approach offers great potential for early diagnosis of ASD. The use of thermal imaging as a passive method to analyze ASD-related physiological signals has been proposed. In previous research, a deep learning model was developed to classify the faces of autistic children using thermal images. Therefore, this study aims to create a new Thermal Image Classification model for Autistic Children Using Res-Net Architecture. The architectures applied are ResNet-18, ResNet-34, and ResNet-50. As a comparison system, several of the same parameter values are used: epoch 100, batch size 2, SGD, Cross-entropy, learning rate 0.001, and momentum 0.9. The study test results show that the results of ResNet-18 are 97.22%, ResNet-34 99.22%, and ResNet-50 99.41%. Based on these results, ResNet-50 has the highest value.

Leveraging explainable AI and large-scale datasets for comprehensive classification of renal histologic types

Recently, as the number of cancer patients has increased, much research is being conducted for efficient treatment, including the use of artificial intelligence in genitourinary pathology. Recent research has focused largely on the classification of renal cell carcinoma subtypes. Nonetheless, the broader categorization of renal tissue into non-neoplastic normal tissue, benign tumor and malignant tumor remains understudied. This gap in research can primarily be attributed to the limited availability of extensive datasets including benign tumor and normal tissue in addition to specific type of renal cell carcinoma, which hampers the ability to conduct comprehensive studies in these broader categories. This research introduces a model aimed at classifying renal tissue into three primary categories: normal (non-neoplastic), benign tumor, and malignant tumor. Utilizing digital pathology while slide images (WSIs) from nephrectomy specimens of 2,535 patients from multiple institutions, the model provides a foundational approach for distinguishing these key tissue types. The study utilized a dataset of 12,223 WSIs comprising 1,300 WSIs of normal tissue, 700 WSIs of benign tumors, and 10,223 WSIs of malignant tumors. Employing the ResNet-18 architecture and a Multiple Instance Learning approach, the model demonstrated high accuracy, with F1-scores of 0.934 (CI: 0.933–0.934) for normal tissue, 0.684 (CI: 0.682–0.687) for benign tumors, and 0.878 (CI: 0.877–0.879) for malignant tumors. The overall performance was also notable, achieving a weighted average F1-score of 0.879 (CI: 0.879–0.880) and a weighted average area under the receiver operating characteristic curve of 0.969 (CI: 0.969–0.969). This model significantly aids in the swift and accurate diagnosis of renal tissue, encompassing non-neoplastic normal tissue, benign tumor, and malignant tumor.

Predicting the risk of type 2 diabetes mellitus (T2DM) emergence in 5 years using mammography images: a comparison study between radiomics and deep learning algorithm.

The prevalence of type 2 diabetes mellitus (T2DM) has been steadily increasing over the years. We aim to predict the occurrence of T2DM using mammography images within 5 years using two different methods and compare their performance. We examined 312 samples, including 110 positive cases (developed T2DM after 5 years) and 202 negative cases (did not develop T2DM) using two different methods. In the first method, a radiomics-based approach, we utilized radiomics features and machine learning (ML) algorithms. The entire breast region was chosen as the region of interest for extracting radiomics features. Then, a binary breast image was created from which we extracted 668 features and analyzed them using various ML algorithms. In the second method, a complex convolutional neural network (CNN) with a modified ResNet architecture and various kernel sizes was applied to raw mammography images for the prediction task. A nested, stratified five-fold cross-validation was done for both parts A and B to compute accuracy, sensitivity, specificity, and area under the receiver operating curve (AUROC). Hyperparameter tuning was also done to enhance the model's performance and reliability. The radiomics approach's light gradient boosting model gave 68.9% accuracy, 30.7% sensitivity, 89.5% specificity, and 0.63 AUROC. The CNN method achieved an AUROC of 0.58 over 20 epochs. Radiomics outperformed CNN by 0.05 in terms of AUROC. This may be due to the more straightforward interpretability and clinical relevance of predefined radiomics features compared with the complex, abstract features learned by CNNs.

Noninvasive Anemia Detection and Hemoglobin Estimation from Retinal Images Using Deep Learning: A Scalable Solution for Resource-Limited Settings.

The purpose of this study was to develop and validate a deep-learning model for noninvasive anemia detection, hemoglobin (Hb) level estimation, and identification of anemia-related retinal features using fundus images. The dataset included 2265 participants aged 40years and above from a population-based study in South India. The dataset included ocular and systemic clinical parameters, dilated retinal fundus images, and hematological data such as complete blood counts and Hb concentration levels. Eighty percent of the dataset was used for algorithm development and 20% for validation. A deep-convolutional neural network, utilizing VGG16, ResNet50, and InceptionV3 architectures, was trained to predict anemia and estimate Hb levels. Sensitivity, specificity, and accuracy were calculated, and receiver operating characteristic (ROC) curves were generated for comparison with clinical anemia data. GradCAM saliency maps highlighted regions linked to anemia and image processing techniques to quantify anemia-related features. For predicting anemia, the InceptionV3 model demonstrated the best performance, achieving 98% accuracy, 99% sensitivity, 97% specificity, and an area under the curve (AUC) of 0.98 (95% confidence interval [CI] = 0.97-0.99). For estimating Hb levels, the mean absolute error for the InceptionV3 model was 0.58 g/dL (95% CI = 0.57-0.59 g/dL). The model focused on the area around the optic disc and the neighboring retinal vessels, revealing that anemic subjects exhibited significantly increased vessel tortuosity and reduced vessel density (P < 0.001), with variable effects on vessel thickness. The InceptionV3 model accurately predicted anemia and Hb levels, highlighting the potential of deep learning and vessel analysis for noninvasive anemia detection. The proposed method offers the possibility to quantitatively predict hematological parameters in a noninvasive manner.

Development of a Clinically Applicable Deep Learning System Based on Sparse Training Data to Accurately Detect Acute Intracranial Hemorrhage from Non-enhanced Head Computed Tomography

Non-enhanced head computed tomography is widely used for patients presenting with head trauma or stroke, given acute intracranial hemorrhage significantly influences clinical decision-making. This study aimed to develop a deep learning algorithm, referred to as DeepCT, to detect acute intracranial hemorrhage on non-enhanced head computed tomography images and evaluate its clinical applicability. We retrospectively collected 1,815 computed tomography image sets from a single center for model training. Additional computed tomography sets from 3 centers were used to construct an independent validation dataset (VAL) and 2 test datasets (GPS-C and DICH). A third test dataset (US-TW) comprised 150 cases, each from 1 hospital in Taiwan and 1 hospital in the United States of America. Our deep learning model, based on U-Net and ResNet architectures, was implemented using PyTorch. The deep learning algorithm exhibited high accuracy across the validation and test datasets, with overall accuracy ranging from 0.9343 to 0.9820. Our findings show that the deep learning algorithm effectively identifies acute intracranial hemorrhage in non-enhanced head computed tomography studies. Clinically, this algorithm can be used for hyperacute triage, reducing reporting times, and enhancing the accuracy of radiologist interpretations. The evaluation of the algorithm on both United States and Taiwan datasets further supports its universal reliability for detecting acute intracranial hemorrhage.

Resnet for Blood Sample Detection: A Study on Improving Diagnostic Accuracy

Automated blood cell analysis plays a crucial role in medical diagnostics, enabling rapid and accurate assessment of a patient's health status. In this paper, we provide a unique technique for detecting and classifying WBCs,RBCs, and platelets inside blood smear pictures using ResNet (Residual Neural Network), a deep learning architecture. Because of its capacity to efficiently train very deep neural networks while minimizing the vanishing gradient problem, the ResNet architecture has exhibited excellent performance in a variety of image recognition applications. Leveraging the power of ResNet, we developed a multi-class classification model capable of distinguishing between WBCs, RBCs, and platelets within microscopic images of blood smears. Our methodology involved preprocessing the blood smear images to enhance contrast and remove noise, followed by image segmentation to isolate individual blood cells and platelets. The segmented images were then used to train and fine-tune a ResNet model, utilizing a large annotated dataset of labeled blood cell images. The trained model exhibited remarkable accuracy in identifying and classifying different blood cell types, even in the presence of overlapping cells or artifacts. We extensively tested our suggested technique, on a range of blood smear images to evaluate its performance. The findings demonstrated that ResNet effectively identifies and categorizes WBCs, (RBCs) and platelets. When compared to methods our approach showcased superior accuracy, robustness and generalization capabilities. After training the model with the Resnet algorithm we got 92% of Accuracy

Deep Learning for Multi-Label Facial Attribute Classification on Large-Scale Image Datasets (CelebA)

The exponential growth of data in recent years has led to an increasing demand for advanced techniques, especially those that work on large and complex data. This has given deep learning a significant advance in dealing with the tasks of analyzing, improving, and distinguishing big data. Our research focused on CNNs from this data and applying deep learning algorithms and their analysis to a large-scale image dataset. More specifically, our research focused on a dataset called CelebA, which contains more than 200,000 face images annotated with 40 binary facial features. It is a multi-label classification model based on the ResNet-50 architecture that has been fine-tuned to predict different facial features and hair color such as age, gender, and facial expressions. It was also trained using data augmentation, taking into account pose differences and background clutter to reduce imbalance between classes. These results reflect very strong predictive performance, with an average mean accuracy of 0.86 and an overall F1 score of 0.81 across all features. Attributes identified by clear visual cues—for example, “smiling,” “male ”and“ wearing lipstick”—were highly accurate, while less obvious attributes such as “big lips” and “narrow eyes” were more difficult to classify. We would like to point out that the results demonstrate the high efficiency of using deep learning models for multi-label classification on big data while solving problems associated with class imbalance and overfitting models. This research leads to the larger general field of big data analytics; in particular, it demonstrates how deep learning can be efficiently applied to large image datasets for automatic attribute recognition. It also opens up potential applications in areas such as biometric identification, surveillance, and human-computer interaction.

Enhancing Epstein–Barr virus detection in IBD patients with XAI and clinical data integration

BackgroundThere is increasing evidence of a link between Epstein–Barr virus (EBV) and gastrointestinal cancers, particularly in immunocompromised patients and those on thiopurines. Traditional endoscopic techniques have limitations in diagnosing EBV, whereas standard methods for diagnosing EBV have practical challenges in regular clinical settings related to their invasiveness and high costs. MethodsAn explainable AI (XAI) model was developed to analyze 1598 endoscopic images from 287 patients with inflammatory bowel disease (IBD), with a focus on detecting EBV infection. Following the application of data augmentation and transfer learning techniques, the model accurately classified the presence of EBV, and its performance was quantified through receiver operating characteristic (ROC) curve analysis, including calculation of the area under the curve (AUC). Furthermore, a combination of the Grad-CAM method and clinical data analysis significantly improved the interpretability and diagnostic accuracy of the model. ResultsOur model significantly outperformed other models, including those developed with ResNet50, Vision Transformer (ViT), and MobileNet v2 architectures, in detecting EBV on endoscopic images, achieving an accuracy of 73.83 % and an F1-score of 73.70 %. The model showed good performance in distinguishing between EBV-positive and EBV-negative images according to the confusion matrix and ROC curve analysis, with metrics including a true negative rate of 79.76 %, a true positive rate of 67.32 %, and an AUC of 0.74. Additionally, the generated salience maps effectively identified key regions in the images, enhancing lesion detection in patients with EBV infection. Our correlational analysis revealed significant associations between EBV infection and clinical parameters such as age, illness duration, and total bilirubin, and suggested that EBV infection had a notably greater incidence in ulcerative colitis (UC) patients than in Crohn's disease (CD) patients. ConclusionOur study successfully created an XAI-assisted system that allows the accurate detection of EBV in endoscopic images and improves the diagnosis of EBV infection through the integration of clinical data.

AI-enhanced interface for colonic polyp segmentation using DeepLabv3+ with comparative backbone analysis

Polyps are one of the early stages of colon cancer. The detection of polyps by segmentation and their removal by surgical intervention is of great importance for making treatment decisions. Although the detection of polyps through colonoscopy images can lead to multiple expert needs and time losses, it can also include human error. Therefore, automatic, fast, and highly accurate segmentation of polyps from colonoscopy images is important. Many methods have been proposed, including deep learning-based approaches. In this study, a method using DeepLabv3+ with encoder-decoder structure and ResNet architecture as backbone network is proposed for the segmentation of colonic polyps. The Kvasir-SEG polyp dataset was used to train and test the proposed method. After images were preprocessed, the training of the proposed network was performed. The trained network was then tested and performance metrics were calculated, and additionally, a GUI (Graphical User Interface) was designed to enable the segmentation of colonoscopy images for polyp segmentation. The experimental results showed that the ResNet-50 based DeepLabv3+ model had high performance metrics such as DSC: 0.9609, mIoU: 0.9246, demonstrating its effectiveness in the segmentation of colonic polyps. In conclusion, our method utilizing DeepLabv3+ with a ResNet-50 backbone achieves highly accurate colonic polyp segmentation. The obtained results demonstrate its potential to significantly enhance colorectal cancer diagnosis and planning for polypectomy surgery through automated image analysis.&#xD.

Color-based Classification of Dried Cocoa Beans from Various Origins of Indonesia by Image Analysis Using AlexNet and ResNet Architecture-Convolutional Neural Networks

Cocoa plant is widely cultivated in Indonesia and spread across various regions. Diversity in geographical conditions has been known to significantly affect the quality of cocoa beans. Practically, cocoa beans are often mixed without considering the variation in the quality and its origin. This resulted in reduced global quality and product inconsistency. Improved recognition and classification methods are needed to solve those problems. Non-destructive classification methods can be used to provide a more efficient classification process. The use of artificial intelligence with computer-based deep learning methods was used in this study. Beans samples of various origins (Aceh, Bali, Banten, Yogyakarta, East Kalimantan, West Sulawesi, and West Sumatera) were evaluated. From thecollected samples, 9100 images were then taken for data processing. Data preprocessing included denoising of the background image, cropping, resizing andchanging the storage extension through the training-validation stage and the testing process. AlexNet and ResNet architectures on a Convolutional NeuralNetwork were used for classification. The results showed that the average accuracy of cocoa image classification based on color identification by computer machines using Alexnet and ResNet was high (99.91% and 99.99%, respectively). This method can be applied to provide more efficient color-based cocoa bean classification for industrial purposes.

Comparison of Deep Learning and Yolov8 Models for Fox Detection Around the Henhouse

Human beings, who have been engaged in agriculture and animal husbandry for centuries, have to constantly track, take care and maintain their own agricultural lands and animals. This requires constant labor and time. The aim and originality of this study is to identify foxes that kidnap animals such as chickens, geese, ducks and turkeys that live in the coops of individuals engaged in poultry farming. In this way, even if the farmer is not in the henhouse at that moment, material and moral losses to the farmers will be prevented. To achieve this purpose, many images were collected to form dataset. The collected dataset was classified according to whether the fox was in the henhouse or not. Then, the outputs of DenseNet, MobileNet, ResNet50, VGG16, VGG19, Xception and Yolov8 architectures were fine tuned to be performed in transfer learning to detect existence of a fox in the henhouse. Then, the models were trained, and their performances were compared in terms of performance metrics such as loss, accuracy, precision and F1. In the results, Yolov8 architectures generally have demonstrated the best performances.

Pioneering automation in agricultural subsidy processing through deep learning for computer vision

ABSTRACT Validating farmers’ eligibility for subsidies under the European Union’s Common Agricultural Policy (CAP) requires substantial resources. An automated solution can minimize human errors and save time. This study leverages deep learning to automate the validation process using ground-level landscape images (GLLP), an approach not previously explored. We conducted two rigorous experiments to identify the optimal model architecture and assess the impact of data cleaning, accurate labelling, reduced feature ambiguity, and stricter photo collection protocols on performance. Our results reveal that while the ResNet-18 architecture achieved a recall of 0.61 on the original dataset, distinguishing between eligible and ineligible grasslands was challenging. By addressing image quality and enforcing standardized data collection, accuracy improved to 0.89 and recall to 0.94. Crucially, our research emphasizes the development and validation of effective data collection strategies, ensuring uniformity and high-quality data essential for accurate model performance. Our findings affirm the potential of deep learning to automate subsidy eligibility classification, contingent on resolving challenges related to the quality and standardization of GLLP data.

Primary angle-closed diseases recognition through artificial intelligence-based anterior segment-optical coherence tomography imaging.

In this study, artificial intelligence (AI) was used to deeply learn the classification of the anterior segment-Optical Coherence Tomography (AS-OCT) images. This AI systems automatically analyzed the angular structure of the AS-OCT images and automatically classified anterior chamber angle. It would improve the efficiency of AS-OCT image analysis. The subjects were from the glaucoma disease screening and prevention project for elderly people in Shanghai community. Each scan contained 72 cross-sectional AS-OCT frames. We developed a deep learning-based AS-OCT image automatic anterior chamber angle analysis software. Classifier performance was evaluated against glaucoma experts' grading of AS-OCT images as standard. Outcome evaluation included accuracy (ACC) and area under the receiver operator curve (AUC). 94895 AS-OCT images were collected from 687 participants, in which 69,243 images were annotated as open, 16,433 images were annotated as closed, and 9219 images were annotated as non-gradable. The class-balanced train data were formed from randomly extracting the same number of open angle images as the closed angle images, which contained 22,393 images (11127 open, 11256 closed). The best-performing classifier was developed by applying transfer learning to the ResNet-50 architecture. against experts' grading, this classifier achieved an AUC of 0.9635. Deep learning classifiers effectively detect angle closure based on automated analysis of AS-OCT images. This system could be used to automate clinical evaluations of the anterior chamber angle and improve efficiency of interpreting AS-OCT images. The results demonstrated the potential of the deep learning system for rapid recognition of high-risk populations of PACD.

A Fast Coding Unit Partitioning Decision Algorithm for Versatile Video Coding Based on Gradient Feedback Hierarchical Convolutional Neural Network and Light Gradient Boosting Machine Decision Tree

Video encoding technology is a foundational component in the advancement of modern technological applications. The latest standard in universal video coding, H.266/VVC, features a quad-tree with nested multi-type tree (QTMT) partitioning structure, which represents an improvement over its predecessor, High-Efficiency Video Coding (H.265/HEVC). This configuration facilitates adaptable block segmentation, albeit at the cost of heightened encoding complexity. In view of the aforementioned considerations, this paper puts forth a deep learning-based approach to facilitate CU partitioning, with the aim of supplanting the intricate CU partitioning process observed in the Versatile Video Coding Test Model (VTM). We begin by presenting the Gradient Feedback Hierarchical CNN (GFH-CNN) model, an advanced convolutional neural network derived from the ResNet architecture, enabling the extraction of features from 64 × 64 coding unit (CU) blocks. Following this, a hierarchical network diagram (HND) is crafted to depict the delineation of partition boundaries corresponding to the various levels of the CU block’s layered structure. This diagram maps the features extracted by the GFH-CNN model to the partitioning at each level and boundary. In conclusion, a LightGBM-based decision tree classification model (L-DT) is constructed to predict the corresponding partition structure based on the prediction vector output from the GFH-CNN model. Subsequently, any errors in the partitioning results are corrected in accordance with the encoding constraints specified by the VTM, which ultimately determines the final CU block partitioning. The experimental results demonstrate that, in comparison with VTM-10.0, the proposed algorithm achieves a 48.14% reduction in complexity with only a 0.83% increase in bitrate under the top-three configuration, which is negligible. In comparison, the top-two configuration resulted in a higher complexity reduction of 63.78%, although this was accompanied by a 2.08% increase in bitrate. These results demonstrate that, in comparison to existing solutions, our approach provides an optimal balance between encoding efficiency and computational complexity.