Convolutional Neural Network Training Research Articles

VeriCNN: Integrity verification of large-scale CNN training process based on zk-SNARK

Cloud environments are frequently employed to train Convolutional Neural Networks (CNNs). On the other hand, the training procedure is exposed to potential integrity threats due to the cloud’s lack of trustworthiness. Although there are now available methods for zero-knowledge proof-based training integrity verification for small models, the low-proof performance of large-scale CNNs like LeNet-5 and VGG16 makes this task challenging. To solve the training integrity verification of large-scale CNNs, this research suggests a technique called VeriCNN. Particularly, VeriCNN builds proof of integrity for CNN training using zk-SNARK. It adopts the Winograd algorithm to optimize convolution operations and introduces a high-probability matrix multiplication to optimize zero-knowledge proofs. Experimental results demonstrate the effectiveness and practicality of VeriCNN in verifying the training integrity of large-scale models. For instance, VeriCNN can generate a training integrity proof for the LeNet-5 model in just 25 s and for the VGG16 model in 390 s.

PATH-13. LEARNED RESIZING WITH EFFICIENT TRAINING (LRET) FACILITATES IMPROVED PERFORMANCE OF LARGE-SCALE BRAIN TUMOR HISTOLOGY IMAGE CLASSIFICATION MODELS

Abstract BACKGROUND Histologic examination is vital in oncology research and diagnostics. The adoption of digital scanning of whole slide images (WSI) has created an opportunity to leverage deep learning-based image classification methods to enhance diagnosis and risk stratification. However, technical limitations prevent training and deployment of accurate comprehensive multiclass deep convolutional neural networks (DCNN) models for histopathology image classification. The input dimensions of DCNN architectures are small compared to the typical pathologist field of view, degrading performance by excluding important architectural features. Furthermore, data requirements for comprehensive models are sufficiently large to overwhelm the system memory during training. METHODS A method termed Learned Resizing with Efficient Training (LRET) was developed to address the main limitations of traditional histopathology classification model training. The LRET method couples efficient training techniques with image resizing to facilitate seamless integration of larger histology image patches into state-of-the-art classification models while preserving important structural information. The LRET method was coupled with two distinct resizing techniques to train three diverse histology image datasets using five different DCNN architectures. Performance metrics were compared on cross validation and hold out test sets. RESULTS LRET-trained models were flexible to multiple input patch dimensions and DCNN models. We demonstrated performance improvement across all datasets while significantly reducing the training time and resources over traditional methods. Using a large-scale, multiclass brain tumor classification dataset consisting of 74 distinct histopathologic classes, LRET-trained models outperformed existing methods by 15-28% in accuracy, yielding 94% accuracy for the best model. CONCLUSION The LRET method for DCNN training significantly enhances the performance of large-scale multiclass histopathology image classification. The implications of this work extend to broader applications within medical imaging and beyond, where efficient integration of high-resolution images into deep learning pipelines is paramount for driving advancements research and clinical practice.

Prediction of photovoltaic power generation based on a hybrid model

In order to fully exploit the relationship between temporal features in photovoltaic power generation data and improve the prediction accuracy of photovoltaic power generation, a photovoltaic power generation forecasting method is proposed based on a hybrid model of the convolutional neural network (CNN) and extreme gradient boost (XGBoost). Taking the historical data of China’s photovoltaic power plants as a sample, the high-dimensional mapping relationship of photovoltaic power generation variables is extracted based on the convolutional layer and pooling layer of the CNN network to construct a high-dimensional time-series feature vector, which is an input for the XGBoost. A photovoltaic power generation prediction model is established based on CNN-XGBoost by training CNN and XGBoost parameters. Since it is difficult for a single model to achieve optimal prediction accuracy under different weather conditions, the k-means clustering algorithm is used to group the power datasets and train independent models to improve prediction accuracy. Through the actual data verification of photovoltaic power plants, the proposed photovoltaic power generation prediction model can accurately predict the power, which shows high prediction accuracy and generalization ability compared with other methods.

Cross-domain prototype similarity correction for few-shot radar modulation signal recognition

The new classes of radar signals are increasingly difficult to acquire under non-cooperative environments, which makes it difficult to support convolutional neural network training with limited labeled samples. The few-shot learning (FSL) methods have shown great performance in classification with limited labeled samples, but the FSL methods ignore that the class distributions between the new and original tasks are significantly different, resulting in a massive challenge in identifying new radar signals. To solve this problem, a few-shot radar modulation signal recognition method based on cross-domain prototype similarity correction (CDPSC) is proposed. Specifically, a residual feature tokenizer transformer (RFTT) model embedded with a pooling token generation block is designed to focus on the important features and improve the ability to represent samples. Meanwhile, the proposed domain prototype similarity mapping (DPSM) strategy adaptively learns the class mapping, reduces the inter-domain difference through feature distribution alignment, and effectively corrects the target domain prototypes. In addition, we introduce a sample prototype embedding (SPE) strategy in the training phase, which can reduce the intra-class distance and increase the inter-class distance. Experimental results demonstrate that the CDPSC method is superior to typical FSL methods in recognition accuracy under different sample numbers.

ASD-SWNet: a novel shared-weight feature extraction and classification network for autism spectrum disorder diagnosis

The traditional diagnostic process for autism spectrum disorder (ASD) is subjective, where early and accurate diagnosis significantly affects treatment outcomes and life quality. Thus, improving ASD diagnostic methods is critical. This paper proposes ASD-SWNet, a new shared-weight feature extraction and classification network. It resolves the issue found in previous studies of inefficiently integrating unsupervised and supervised learning, thereby enhancing diagnostic precision. The approach utilizes functional magnetic resonance imaging to improve diagnostic accuracy, featuring an autoencoder (AE) with Gaussian noise for robust feature extraction and a tailored convolutional neural network (CNN) for classification. The shared-weight mechanism utilizes features learned by the AE to initialize the convolutional layer weights of the CNN, thereby integrating AE and CNN for joint training. A novel data augmentation strategy for time-series medical data is also introduced, tackling the problem of small sample sizes. Tested on the ABIDE-I dataset through nested ten-fold cross-validation, the method achieved an accuracy of 76.52% and an AUC of 0.81. This approach surpasses existing methods, showing significant enhancements in diagnostic accuracy and robustness. The contribution of this paper lies not only in proposing new methods for ASD diagnosis but also in offering new approaches for other neurological brain diseases.

A Predictive Model for Intraoperative Cerebrospinal Fluid Leak During Endonasal Pituitary Adenoma Resection Using a Convolutional Neural Network

Cerebrospinal fluid (CSF) leak during endoscopic endonasal transsphenoidal surgery can lead to postoperative complications. The clinical and anatomic risk factors of intraoperative CSF leak are not well defined. We applied a two-dimensional (2D) convolutional neural network (CNN) machine learning model to identify risk factors from preoperative magnetic resonance imaging. All adults who underwent endoscopic endonasal transsphenoidal surgery at our institution from January 2007 to March 2023 who had accessible preoperative stereotactic magnetic resonance imaging were included. A retrospective classic statistical analysis was performed to identify demographic, clinical, and anatomic risk factors of intraoperative CSF leak. Stereotactic T2-weighted brain magnetic resonance imaging scans were used to train and test a 2D CNN model. Of 220 included patients, 81 (36.8%) experienced intraoperative CSF leak. Among all preoperative variables, visual disturbance was the only statistically significant identified risk factor (P= 0.008). The trained 2D CNN model predicted CSF leak with 92% accuracy and area under receiver operating characteristic curve of 0.90 (sensitivity of 86% and specificity of 93%). Class activation mapping of this model revealed that anatomic regions of CSF flow were most important in predicting CSF leak. Further review of the class activation mapping gradients revealed regions of the diaphragma sellae, clinoid processes, temporal horns, and optic nerves to have anatomic correlation to intraoperative CSF leak risk. Additionally, visual disturbances from anatomic compression of the optic chiasm were the only identified clinical risk factor. Our 2D CNN model can help a treating team to better anticipate and prepare for intraoperative CSF leak.

Deep learning and saliency-based parking IoT classification under different weather conditions

The world’s population is growing at an exponential rate, which has created a slew of new issues in our daily lives. Among these challenges is the difficulty of finding available parking spaces. To overcome this obstacle, an effective system for detecting vacant parking spaces is crucial, particularly in densely populated areas. Even though there are numerous parking detecting systems on the market today, they all have their own set of restrictions. There are primarily two types of parking systems available. The first is a hardware-based system that uses hardware to detect whether a space is vacant or occupied, and the second is a smart parking system that uses cameras to classify parking status. Vision-based parking systems are becoming increasingly popular due to benefits such as flexibility and cost efficiency. However, due to challenging weather conditions and various occlusions caused by objects, such systems are challenging to develop. The limitations of a vision-based parking system must be addressed for it to become more accurate. To date, many researchers have attempted to incorporate deep learning-based parking systems with various algorithms but none of them has focused on reducing the complexity of Convolutional Neural Network (CNN) training through unsupervised pre-processing steps. In this study, an attempt is made to reduce the complexity of CNN by integrating salient features as a pre-processing step for normalizing weather conditions. It has been observed that salient features-based pre-processing makes the system more robust in terms of accuracy and time efficiency.

Training of Convolutional Neural Networks for Image Classification with Fully Decoupled Extended Kalman Filter

Training of Convolutional Neural Networks for Image Classification with Fully Decoupled Extended Kalman Filter

Scoring facial attractiveness with deep convolutional neural networks: How training on standardized images reduces the bias of facial expressions.

In many medical disciplines, facial attractiveness is part of the diagnosis, yet its scoring might be confounded by facial expressions. The intent was to apply deep convolutional neural networks (CNN) to identify how facial expressions affect facial attractiveness and to explore whether a dedicated training of the CNN is able to reduce the bias of facial expressions. Frontal facial images (n = 840) of 40 female participants (mean age 24.5 years) were taken adapting a neutral facial expression and the six universal facial expressions. Facial attractiveness was computed by means of a face detector, deep convolutional neural networks, standard support vector regression for facial beauty, visual regularized collaborative filtering and a regression technique for handling visual queries without rating history. CNN was first trained on random facial photographs from a dating website and then further trained on the Chicago Face Database (CFD) to increase its suitability to medical conditions. Both algorithms scored every image for attractiveness. Facial expressions affect facial attractiveness scores significantly. Scores from CNN additionally trained on CFD had less variability between the expressions (range 54.3-60.9 compared to range: 32.6-49.5) and less variance within the scores (P ≤ .05), but also caused a shift in the ranking of the expressions' facial attractiveness. Facial expressions confound attractiveness scores. Training on norming images generated scores less susceptible to distortion, but more difficult to interpret. Scoring facial attractiveness based on CNN seems promising, but AI solutions must be developed on CNN trained to recognize facial expressions as distractors.

Deep learning-based classification of erosion, synovitis and osteitis in hand MRI of patients with inflammatory arthritis

ObjectivesTo train, test and validate the performance of a convolutional neural network (CNN)-based approach for the automated assessment of bone erosions, osteitis and synovitis in hand MRI of patients with...

Soil Surface Texture Classification Using RGB Images

The classification of soil surface texture is critical for agricultural and environmental management, impacting properties like water-holding capacity and nutrient retention. Traditional methods relying on controlled environments face scalability challenges. This study introduces a scalable and high-resolution classification method using RGB images captured under uncontrolled field conditions (UFC). The proposed approach integrates image processing techniques, texture-enhancing filters, and Convolutional Neural Networks (CNNs) for accurate soil texture classification. The method involves image segmentation, tiling, application of Gabor filters, and CNN training, demonstrating improved performance over traditional techniques. Keywords: Soil texture classification, RGB images, uncontrolled field conditions, image processing, Gabor filters, convolutional neural networks.

FLsM: Fuzzy Localization of Image Scenes Based on Large Models

This article primarily focuses on the study of image-based localization technology. While traditional methods have made significant advancements in technology and applications, the emerging field of visual image-based localization technology demonstrates tremendous potential for research. Deep learning has exhibited a strong performance in image processing, particularly in developing visual navigation and localization techniques using large-scale visual models. This paper introduces a sophisticated scene image localization technique based on large models in a vast spatial sample environment. The study involved training convolutional neural networks using millions of geographically labeled images, extracting image position information using large model algorithms, and collecting sample data under various conditions in elastic scene space. Through visual computation, the shooting position of photos was inferred to obtain the approximate position information of users. This method utilizes geographic location information to classify images and combines it with landmarks, natural features, and architectural styles to determine their locations. The experimental results show variations in positioning accuracy among different models, with the most optimal model obtained through training on a large-scale dataset. They also indicate that the positioning error in urban street-based images is relatively small, whereas the positioning effect in outdoor and local scenes, especially in large-scale spatial environments, is limited. This suggests that the location information of users can be effectively determined through the utilization of geographic data, to classify images and incorporate landmarks, natural features, and architectural styles. The study’s experimentation indicates the variation in positioning accuracy among different models, highlighting the significance of training on a large-scale dataset for optimal results. Furthermore, it highlights the contrasting impact on urban street-based images versus outdoor and local scenes in large-scale spatial environments.

Exploring 2D X-ray diffraction phase fraction analysis with convolutional neural networks: Insights from kinematic-diffraction simulations

Deep-learning models are effective for analyzing the complex information in 2D X-ray diffraction (XRD) patterns. Accurately collecting parameters of the material sample is crucial during model training, significantly impacting model performance. In this study, we employ a kinematic-diffraction simulator to generate simulated 2D XRD patterns for Ti–6Al–4V alloy, allowing precise control of sample parameters. These simulated patterns are used to train convolutional neural networks, predicting β\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\upbeta$$\\end{document}-phase volume fractions. The training data set consists exclusively of 2D XRD patterns with pure α\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\upalpha$$\\end{document}- or pure β\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\upbeta$$\\end{document}-phase, while the testing set incorporates patterns with intermediate phase volume fraction. In particular, we investigate how the architectures of the model influence prediction reliability and computational performance. Experimental results reveal that, with appropriate training, the convolutional neural network accurately detects intermediate phase volume fractions even trained with only pure-phase patterns, achieving a mean square error accuracy of 9.4×10-4\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$9.4 \ imes 10^{-4}$$\\end{document}.Graphical abstract

MResTNet: A Multi-Resolution Transformer Framework with CNN Extensions for Semantic Segmentation.

A fundamental task in computer vision is the process of differentiation and identification of different objects or entities in a visual scene using semantic segmentation methods. The advancement of transformer networks has surpassed traditional convolutional neural network (CNN) architectures in terms of segmentation performance. The continuous pursuit of optimal performance, with respect to the popular evaluation metric results, has led to very large architectures that require a significant amount of computational power to operate, making them prohibitive for real-time applications, including autonomous driving. In this paper, we propose a model that leverages a visual transformer encoder with a parallel twin decoder, consisting of a visual transformer decoder and a CNN decoder with multi-resolution connections working in parallel. The two decoders are merged with the aid of two trainable CNN blocks, the fuser that combined the information from the two decoders and the scaler that scales the contribution of each decoder. The proposed model achieves state-of-the-art performance on the Cityscapes and ADE20K datasets, maintaining a low-complexity network that can be used in real-time applications.

Application of Convolutional Networks for Localization and Prediction of Scalar Parameters of Fractured Geological Inclusion

Seismic inversion is an important part of the modern geological exploration process. Novel applications of deep learning are capable of handling heterogeneous media, but require too much data for training. In this paper, we focus on the prediction of fracture inclusion location and its parameters in rock media and approach the problem in the multi-task manner. For this, several multi-task convolutional neural network (CNN) architectures are proposed and compared. The direct seismic problem is considered in the heterogeneous fractured geological model based on the well-known Marmousi2 model in a two-dimensional case. The model of the deformable solid medium containing slip planes with nonlinear slip conditions on them and explicit–implicit numerical method is applied to obtain the synthetic seismic dataset for CNN training and validation.

Student Progression and Dropout Rates Using Convolutional Neural Network: A Case Study of the Arab Open University

Pre-trained convolutional neural network (CNN) structures are considered as one of the emerging education management tools that can help improve the quality of education by allowing decision makers to manipulate important indicators. These indicators, which are categorized as student and institution specific factors, may influence student progress, retention or dropout rates. In this paper, we develop a deep learning model of predicting students’ satisfactions and their expected outcomes and associated early failures. The model can also predict dropout rates and identify the main baseline risk factors that influence such rates. The academic data of 12,000 students enrolled from 2018 in the Arab Open University student information system are used as CNNs training dataset to ensure that all institution levels are represented. Then, the trained network provides a probabilistic model that indicates, for each student, the probability of dropout. Based on the prediction model, the study presents an early warning system framework to generate alerts and recommendations to allow early and effective institutional intervention. Experiments are achieved by using the proposed dataset and the performance of our approach is considerably better compared to the competitive models in terms of training/validation accuracy and mean square errors.

Quality Evaluation of Road Surface Markings with Uncertainty Aware Regression and Progressive Pretraining

Maintaining high-quality road markings is essential for both safety and traffic flow. However, there has been limited research on automating the process of evaluating the quality of these markings and identifying degraded ones that need to be fixed. Our paper introduces a new approach that uses uncertainty aware (UA) regression to evaluate the quality of road surface markings. The approach is based on deep learning models and a unique training method called “progressive pretraining (PPT).” We used a dataset of RGB images which we converted to binary masks. These masks were then augmented and used to train convolutional neural networks models with a PPT strategy. The results showed that both the hybrid and UA models managed to outperform the baseline model in some metrics such as mean average error which was at 24.38% and accuracy with 81.27%. Additionally, each model showed unique strengths across various performance metrics, highlighting the efficacy of integrating uncertainty and progressive learning in quality assessment tasks. This study presents a solid proof of concept for the application of UA methods in quality evaluation tasks in general, and surface marking quality evaluation in particular.

The Impact of Drop Test Conditions on Brain Strain Location and Severity: A Novel Approach Using a Deep Learning Model

In contact sports such as rugby, players are at risk of sustaining traumatic brain injuries (TBI) due to high-intensity head impacts that generate high linear and rotational accelerations of the head. Previous studies have established a clear link between high-intensity head impacts and brain strains that result in concussions. This study presents a novel approach to investigating the effect of a range of laboratory controlled drop test parameters on regional peak and mean maximum principal strain (MPS) predictions within the brain using a trained convolutional neural network (CNN). The CNN is publicly available at https://github.com/Jilab-biomechanics/CNN-brain-strains. The results of this study corroborate previous findings that impacts to the side of the head result in significantly higher regional MPS than forehead impacts. Forehead impacts tend to result in the lowest region-averaged MPS values for impacts where the surface angle was at 0° and 45°, while side impacts tend to result in higher regional peak and mean MPS. The absence of a neck in drop tests resulted in lower regional peak and mean MPS values. The results indicated that the relationship between drop test parameters and resulting regional peak and mean MPS predictions is complex. The study’s findings offer valuable insights into how deep learning models can be used to provide more detailed insights into how drop test conditions impact regional MPS. The novel approach used in this paper to predict brain strains can be applied in the development of better methods to reduce the brain strain resulting from head accelerations such as protective sports headgear.

Web-based diagnostic platform for microorganism-induced deterioration on paper-based cultural relics with iterative training from human feedback

PurposePaper-based artifacts hold significant cultural and social values. However, paper is intrinsically fragile to microorganisms, such as mold, due to its cellulose composition, which can serve as a microorganisms’ nutrient source. Mold not only can damage papers’ structural integrity and pose significant challenges to conservation works but also may subject individuals attending the contaminated artifacts to health risks. Current approaches for strain identification usually require extensive training, prolonged time for analysis, expensive operation costs, and higher risks of secondary damage due to sampling. Thus, in current conservation practices with mold-contaminated artifacts, little pre-screening or strain identification was performed before mold removal, and the cleaning techniques are usually broad-spectrum rather than strain-specific. With deep learning showing promising applications across various domains, this study investigated the feasibility of using a convolutional neural network (CNN) for fast in-situ recognition and classification of mold on paper.MethodsMolds were first non-invasively sampled from ancient Xuan Paper-based Chinese books from the Qing and Ming dynasties. Strains were identified using molecular biology methods and the four most prevalent strains were inoculated on Xuan paper to create mockups for image collection. Microscopic images of the molds as well as their stains situated on paper were collected using a compound microscope and commercial microscope lens for cell phone cameras, which were then used for training CNN models with a transfer learning scheme to perform the classification of mold. To enable involvement and contribution from the research community, a web interface that actuates the process while providing interactive features for users to learn about the information of the classified strain was constructed. Moreover, a feedback functionality in the web interface was embedded for catching potential classification errors, adding additional training images, or introducing new strains, all to refine the generalizability and robustness of the model.Results & ConclusionIn the study, we have constructed a suite of high-confidence classification CNN models for the diagnostic process for mold contamination in conservation. At the same time, a web interface was constructed that allows recurrently refining the model with human feedback through engaging the research community. Overall, the proposed framework opens new avenues for effective and timely identification of mold, thus enabling proactive and targeted mold remediation strategies in conservation.

Machine learning for non-experts: A more accessible and simpler approach to automatic benthic habitat classification

Automating identification of benthic habitats from imagery, with Machine Learning (ML), is necessary to contribute efficiently and effectively to marine spatial planning. A promising method is to adapt pre-trained general convolutional neural networks (CNNs) to a new classification task (transfer learning). However, this is often inaccessible to a non-specialist, requiring large investments in computational resources and time (for user comprehension and model training). In this paper, we demonstrate a simpler transfer learning framework for classifying broad deep-sea benthic habitats. Specifically, we take an ‘off-the-shelf’ CNN (VGG16) and use it to extract features (pixel patterns) from benthic images (without further training). The default outputs of VGG16 are then fed in to a Support Vector Machine (SVM), a classical and simpler method than deep networks. For comparison, we also train the remaining classification layers of VGG16 using stochastic gradient descent. The discriminative power of these approaches is demonstrated on three benthic datasets (574–8353 images) from Norwegian waters; each using a unique imaging platform. Benthic habitats are broadly classified as Soft Substrate (sands, muds), Hard Substrate (gravels, cobbles and boulders) and Reef (Desmophyllum pertusum). We found that the relatively simplicity of the SVM classifier did not compromise performance. Results were competitive with the CNN classifier and consistently high, with test accuracy ranging from 0.87 to 0.95 (average = 0.9 (±0.04)) across datasets, somewhat increasing with dataset size. Impressively, these results were achieved 2.4–5× faster than CNN training and had significantly less dependency on high-specification hardware. Our suggested approach maximises conceptual and practical simplicity, representing a realistic baseline for novice users when approaching benthic habitat classification. This method has wide potential. It allows automated image grouping to aid annotation or further model selection, as well as screening of old-datasets. It is especially suited to offshore scenarios as it can provide quick, albeit crude, insights into habitat presence, allowing adaptation of sampling protocols in near real-time.