Training Deep Learning Models Research Articles

ODNet: A High Real-Time Network Using Orthogonal Decomposition for Few-Shot Strip Steel Surface Defect Classification.

Strip steel plays a crucial role in modern industrial production, where enhancing the accuracy and real-time capabilities of surface defect classification is essential. However, acquiring and annotating defect samples for training deep learning models are challenging, further complicated by the presence of redundant information in these samples. These issues hinder the classification of strip steel surface defects. To address these challenges, this paper introduces a high real-time network, ODNet (Orthogonal Decomposition Network), designed for few-shot strip steel surface defect classification. ODNet utilizes ResNet as its backbone and incorporates orthogonal decomposition technology to reduce the feature redundancies. Furthermore, it integrates skip connection to preserve essential correlation information in the samples, preventing excessive elimination. The model optimizes the parameter efficiency by employing Euclidean distance as the classifier. The orthogonal decomposition not only helps reduce redundant image information but also ensures compatibility with the Euclidean distance requirement for orthogonal input. Extensive experiments conducted on the FSC-20 benchmark demonstrate that ODNet achieves superior real-time performance, accuracy, and generalization compared to alternative methods, effectively addressing the challenges of few-shot strip steel surface defect classification.

Transfer learning with cross-domain adaptation of vibro-acoustic signals for electric motor mechanical fault detection and diagnosis.

Contact-based vibration measurement techniques involving accelerometers are typically utilized for machine health monitoring and manufacturing end-of-line quality control. Acoustic signals, encompassing sound pressure and particle velocity, carry crucial information about the operational status of underlying equipment and can be utilized for detecting various machine faults. Utilizing one-dimensional Convolution Neural Networks (1D-CNN) based on deep learning for processing raw acoustic time signals has proven to be a valuable approach in detecting machinery faults, offering an effective alternative to using traditional accelerometer signals. However, the success of deep learning methods is heavily dependent on the quantity and availability of data for robust network training. When transitioning to new non-contact type sensor technology, manufacturers may face challenges due to insufficient data for training deep learning models. In such cases, leveraging existing historical accelerometer-based data becomes crucial. Transfer learning emerges as a promising solution, allowing deep learning models trained on a source domain (SD) to be applied to a separate but related target domain (TD). This paper hence introduces a domain adaptation methodology by implementing transfer learning, where the SD 1D-CNN network is trained on accelerometer signals, then the trained model weights are transferred to the TD 1D-CNN network that can process acoustic signals.

Few-Shot Conditional Learning: Automatic and Reliable Device Classification for Medical Test Equipment.

The limited availability of specialized image databases (particularly in hospitals, where tools vary between providers) makes it difficult to train deep learning models. This paper presents a few-shot learning methodology that uses a pre-trained ResNet integrated with an encoder as a backbone to encode conditional shape information for the classification of neonatal resuscitation equipment from less than 100 natural images. The model is also strengthened by incorporating a reliability score, which enriches the prediction with an estimation of classification reliability. The model, whose performance is cross-validated, reached a median accuracy performance of over 99% (and a lower limit of 73.4% for the least accurate model/fold) using only 87 meta-training images. During the test phase on complex natural images, performance was slightly degraded due to a sub-optimal segmentation strategy (FastSAM) required to maintain the real-time inference phase (median accuracy 87.25%). This methodology proves to be excellent for applying complex classification models to contexts (such as neonatal resuscitation) that are not available in public databases. Improvements to the automatic segmentation strategy prior to the extraction of conditional information will allow a natural application in simulation and hospital settings.

Utilizing ChatGPT for Curriculum Learning in Developing a Clinical Grade Pneumothorax Detection Model: A Multisite Validation Study.

Background: Pneumothorax detection is often challenging, particularly when radiographic features are subtle. This study introduces a deep learning model that integrates curriculum learning and ChatGPT to enhance the detection of pneumothorax in chest X-rays. Methods: The model training began with large, easily detectable pneumothoraces, gradually incorporating smaller, more complex cases to prevent performance plateauing. The training dataset comprised 6445 anonymized radiographs, validated across multiple sites, and further tested for generalizability in diverse clinical subgroups. Performance metrics were analyzed using descriptive statistics. Results: The model achieved a sensitivity of 0.97 and a specificity of 0.97, with an area under the curve (AUC) of 0.98, demonstrating a performance comparable to that of many FDA-approved devices. Conclusions: This study suggests that a structured approach to training deep learning models, through curriculum learning and enhanced data extraction via natural language processing, can facilitate and improve the training of AI models for pneumothorax detection.

Probing perfection: The relentless art of meddling for pulmonary airway segmentation from HRCT via a human-AI collaboration based active learning method

In the realm of pulmonary tracheal segmentation, the scarcity of annotated data stands as a prevalent pain point in most medical segmentation endeavors. Concurrently, most Deep Learning (DL) methodologies employed in this domain invariably grapple with other dual challenges: the inherent opacity of ‘black box’ models and the ongoing pursuit of performance enhancement. In response to these intertwined challenges, the core concept of our Human-Computer Interaction (HCI) based learning models (RS_UNet, LC_UNet, UUNet and WD_UNet) hinge on the versatile combination of diverse query strategies and an array of deep learning models. We train four HCI models based on the initial training dataset and sequentially repeat the following steps 1–4: (1) Query Strategy: Our proposed HCI models selects those samples which contribute the most additional representative information when labeled in each iteration of the query strategy (showing the names and sequence numbers of the samples to be annotated). Additionally, in this phase, the model selects the unlabeled samples with the greatest predictive disparity by calculating the Wasserstein Distance, Least Confidence, Entropy Sampling, and Random Sampling. (2) Central line correction: The selected samples in previous stage are then used for domain expert correction of the system-generated tracheal central lines in each training round. (3) Update training dataset: When domain experts are involved in each epoch of the DL model's training iterations, they update the training dataset with greater precision after each epoch, thereby enhancing the trustworthiness of the ‘black box’ DL model and improving the performance of models. (4) Model training: Proposed HCI model is trained using the updated training dataset and an enhanced version of existing UNet.Experimental results validate the effectiveness of this Human-Computer Interaction-based approaches, demonstrating that our proposed WD-UNet, LC-UNet, UUNet, RS-UNet achieve comparable or even superior performance than the state-of-the-art DL models, such as WD-UNet with only 15 %–35 % of the training data, leading to substantial reductions (65 %–85 % reduction of annotation effort) in physician annotation time.

DCNN for Pig Vocalization and Non-Vocalization Classification: Evaluate Model Robustness with New Data.

Since pig vocalization is an important indicator of monitoring pig conditions, pig vocalization detection and recognition using deep learning play a crucial role in the management and welfare of modern pig livestock farming. However, collecting pig sound data for deep learning model training takes time and effort. Acknowledging the challenges of collecting pig sound data for model training, this study introduces a deep convolutional neural network (DCNN) architecture for pig vocalization and non-vocalization classification with a real pig farm dataset. Various audio feature extraction methods were evaluated individually to compare the performance differences, including Mel-frequency cepstral coefficients (MFCC), Mel-spectrogram, Chroma, and Tonnetz. This study proposes a novel feature extraction method called Mixed-MMCT to improve the classification accuracy by integrating MFCC, Mel-spectrogram, Chroma, and Tonnetz features. These feature extraction methods were applied to extract relevant features from the pig sound dataset for input into a deep learning network. For the experiment, three datasets were collected from three actual pig farms: Nias, Gimje, and Jeongeup. Each dataset consists of 4000 WAV files (2000 pig vocalization and 2000 pig non-vocalization) with a duration of three seconds. Various audio data augmentation techniques are utilized in the training set to improve the model performance and generalization, including pitch-shifting, time-shifting, time-stretching, and background-noising. In this study, the performance of the predictive deep learning model was assessed using the k-fold cross-validation (k = 5) technique on each dataset. By conducting rigorous experiments, Mixed-MMCT showed superior accuracy on Nias, Gimje, and Jeongeup, with rates of 99.50%, 99.56%, and 99.67%, respectively. Robustness experiments were performed to prove the effectiveness of the model by using two farm datasets as a training set and a farm as a testing set. The average performance of the Mixed-MMCT in terms of accuracy, precision, recall, and F1-score reached rates of 95.67%, 96.25%, 95.68%, and 95.96%, respectively. All results demonstrate that the proposed Mixed-MMCT feature extraction method outperforms other methods regarding pig vocalization and non-vocalization classification in real pig livestock farming.

BRIDGING DATA GAPS: A FEDERATED LEARNING APPROACH TO HEAT EMISSION PREDICTION IN LASER POWDER BED FUSION

Abstract Deep learning has impacted defect prediction in Additive Manufacturing (AM), which is important to ensure process stability and part quality. However, its success depends on extensive training, requiring large, homogenous datasets–remaining a challenge for the AM industry, particularly for small- and medium-sized enterprises (SMEs). The unique and varied characteristics of AM parts, along with the limited resources of SMEs, hampers data collection, posing difficulties in the independent training of deep learning models. Addressing these concerns requires enabling knowledge sharing from the similarities in the physics of the AM process and defect formation mechanisms while carefully handling privacy concerns. Federated learning (FL) offers a solution to allow collaborative model training across multiple entities without sharing local data. This paper introduces an FL framework to predict section-wise heat emission during Laser Powder Bed Fusion (LPBF), a vital process signature. It incorporates a customized Long Short-Term Memory (LSTM) model for each client, capturing the dynamic AM process's time series properties without sharing sensitive information. Three advanced FL algorithms are integrated–FedAvg, FedProx, and FedAvgM–to aggregate model weights rather than raw datasets. Experiments demonstrate that the FL framework ensures convergence and maintains prediction performance comparable to individually trained models. This work demonstrates the potential of FL-enabled AM modeling and prediction where SMEs can improve their product quality without compromising data privacy.

Monitoring tropical freshwater fish with underwater videography and deep learning

Context The application of deep learning to monitor tropical freshwater fish assemblages and detect potential anthropogenic impacts is poorly understood. Aims This study aimed to compare the results between trained human observers and deep learning, using the fish monitoring program for impact detection at Ranger Uranium Mine as a case study. Methods Fish abundance (MaxN) was measured by trained observers and deep learning. Microsoft’s Azure Custom Vision was used to annotate, label and train deep learning models with fish imagery. PERMANOVA was used to compare method, year and billabong. Key results Deep learning model training on 23 fish taxa resulted in mean average precision, precision and recall of 83.6, 81.3 and 89.1%, respectively. PERMANOVA revealed significant differences between the two methods, but no significant interaction was observed in method, billabong and year. Conclusions These results suggest that the distribution of fish taxa and their relative abundances determined by deep learning and trained observers reflect similar changes between control and exposed billabongs over a 3-year period. Implications The implications of these method-related differences should be carefully considered in the context of impact detection, and further research is required to more accurately characterise small-growing schooling fish species, which were found to contribute significantly to the observed differences.

Generative adversarial networks for handwriting image generation: a review

AbstractHandwriting synthesis, the task of automatically generating realistic images of handwritten text, has gained increasing attention in recent years, both as a challenge in itself, as well as a task that supports handwriting recognition research. The latter task is to synthesize large image datasets that can then be used to train deep learning models to recognize handwritten text without the need for human-provided annotations. While early attempts at developing handwriting generators yielded limited results [1], more recent works involving generative models of deep neural network architectures have been shown able to produce realistic imitations of human handwriting [2–19]. In this review, we focus on one of the most prevalent and successful architectures in the field of handwriting synthesis, the generative adversarial network (GAN). We describe the capabilities, architecture specifics, and performance of the GAN-based models that have been introduced to the literature since 2019 [2–14]. These models can generate random handwriting styles, imitate reference styles, and produce realistic images of arbitrary text that was not in the training lexicon. The generated images have been shown to contribute to improving handwriting recognition results when augmenting the training samples of recognition models with synthetic images. The synthetic images were often hard to expose as non-real, even by human examiners, but also could be implausible or style-limited. The review includes a discussion of the characteristics of the GAN architecture in comparison with other paradigms in the image-generation domain and highlights the remaining challenges for handwriting synthesis.

Accurate Vision-Based PCB Positioning Using Cosine-Convolutional Neural Networks

Object positioning is a fundamental task for assembly and inspection operations in the electronic industry. Traditional methods such as template matching and feature-point detection have been applied for object positioning. They are however computationally expensive and are generally affected by environmental changes. Deep learning models based on Convolutional Neural Network (CNN) have also been used for object positioning. They are computationally very efficient but can be sensible under environmental variations such as illumination and noise. In this article, a Cosine-Convolution operation is proposed with the objective of minimizing the effects of illumination variations. The proposed Cosine-Convolution can substitute any convolutional operation in a CNN-based regressor for positioning tasks. The proposed convolution is based on cosine-measure that normalizes the convolution between an image window and a filter at local scope. The proposed Cosine-Convolutional Neural Network (Cosine-CNN) performs more accurate than the traditional CNN under environment variations, including illumination changes, noise, defocusing and template occlusion. The experimental results reveal that the proposed model can reach mean prediction errors smaller than 1 pixel and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1^\circ$</tex-math> </inline-formula> for object shift and rotation, regardless of the environmental variations. The method is computationally very fast, with an average evaluation time of 1.2-ms for real-time PCB positioning. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —The training of deep learning models based on Convolutional Neural Network (CNN) can be tedious. Image samples with different illumination settings must be provided to have an efficient and robust model, which may not be easy to acquire or synthesize. This study proposes a solution to the challenge of illumination variations in images for the object positioning task in Printed Circuit Boards (PCB). A new Convolutional operation called Cosine-Convolutional that normalizes the traditional Convolution is introduced to overcome the illumination variations in the input images. Experimental results demonstrate that the proposed Cosine-Convolutional operation is less sensitive against lightning variations with respect to the training images.

AstroDLLC: Efficiently Reducing Storage and Transmission Costs for Massive Solar Observation Data via Deep Learning-based Lossless Compression

Effective data compression technology is essential for addressing data storage and transmission needs, especially given the escalating volume and complexity of data generated by contemporary astronomy. In this study, we propose utilizing deep learning-based lossless compression techniques to improve compression efficiency. We begin with a qualitative and quantitative analysis of the temporal and spatial redundancy in solar observation data. Based on this analysis, we introduce a novel deep learning-based framework called AstroDLLC for the lossless compression of astronomical solar images. AstroDLLC first segments high-resolution images into blocks to ensure that deep learning model training does not rely on high-computation power devices. It then addresses the non-normality of the partitioned data through simple reversible computational methods. Finally, it utilizes Bit-swap to train deep learning models that capture redundant features across multiple image frames, thereby enhancing compression efficiency. Comprehensive evaluations using data from the New Vacuum Solar Telescope reveal that AstroDLLC achieves a maximum compression ratio of 3.00 per image, surpassing Gzip, RICE, and other lossless technologies. The performance of AstroDLLC underscores its potential to address data compression challenges in astronomy.

Enhanced diabetic retinopathy detection and classification using fundus images with ResNet50 and CLAHE-GAN

Diabetic retinopathy (DR), a progressive eye disorder, can lead to irreversible vision impairment ranging from no DR to severe DR, necessitating precise identification for early treatment. This study introduces an innovative deep learning (DL) approach, surpassing traditional methods in detecting DR stages. It evaluated two scenarios for training DL models on balanced datasets. The first employed image enhancement via contrast limited adaptive histogram equalization (CLAHE) and a generative adversarial network (GAN), while the second did not involve any image enhancement. Tested on the Asia pacific tele-ophthalmology society 2019 blindness detection (APTOS-2019 BD) dataset, the enhanced model (scenario 1) reached 98% accuracy and a 99% Cohen kappa score (CKS), with the non-enhanced model (scenario 2) achieving 95.4% accuracy and a 90.5% CKS. The combination of CLAHE and GAN, termed CLANG, significantly boosted the model's performance and generalizability. This advancement is pivotal for early DR detection and intervention, offering a new pathway to prevent irreversible vision loss in diabetic patients.

Weakly Supervised Deep Learning in Radiology.

Deep learning (DL) is currently the standard artificial intelligence tool for computer-based image analysis in radiology. Traditionally, DL models have been trained with strongly supervised learning methods. These methods depend on reference standard labels, typically applied manually by experts. In contrast, weakly supervised learning is more scalable. Weak supervision comprises situations in which only a portion of the data are labeled (incomplete supervision), labels refer to a whole region or case as opposed to a precisely delineated image region (inexact supervision), or labels contain errors (inaccurate supervision). In many applications, weak labels are sufficient to train useful models. Thus, weakly supervised learning can unlock a large amount of otherwise unusable data for training DL models. One example of this is using large language models to automatically extract weak labels from free-text radiology reports. Here, we outline the key concepts in weakly supervised learning and provide an overview of applications in radiologic image analysis. With more fundamental and clinical translational work, weakly supervised learning could facilitate the uptake of DL in radiology and research workflows by enabling large-scale image analysis and advancing the development of new DL-based biomarkers.

Lip Reading using Deep Learning

Lip reading, the process of interpreting speech by visually observing the movements of the lips, has emerged as a critical area of research with applications spanning communication aids for the hearing impaired, silent speech interfaces, and enhanced human-computer interaction. This paper reviews recent advancements in lip reading technologies, focusing on the integration of machine learning and computer vision techniques. We explore state-of-the-art methods including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer-based models that have significantly improved the accuracy and robustness of lip reading systems. The study highlights the importance of large annotated datasets, such as LipNet and LRW, which have facilitated the training of deep learning models. Additionally, we examine multimodal approaches that combine visual information with audio signals to enhance performance, especially in noisy environments. Despite substantial progress, challenges remain in addressing speaker variability, low resolution, and real-time processing. Future research directions are discussed, emphasizing the need for more diverse datasets, improved model generalization, and real-world application testing. This comprehensive review underscores the potential of advanced lip reading technologies to revolutionize communication accessibility and human-computer interaction. This paper presents the method for Vision based Lip Reading system that uses convolutional neural network (CNN) with attention-based Long Short-Term Memory (LSTM). The dataset includes video clips pronouncing words sentence. The pretrained CNN is used for extracting features from pre- processed video frames which then are processed for learning temporal characteristics by LSTM. The SoftMax layer of architecture provides the result of lip reading. In the present work experiments are performed with two pre-trained models. The system provides 80% accuracy using Tensorflow and ensemble learning. Keywords— CNN; RNN; LSTM; Tensorflow; lip reading; deep learning

Moiré fringe analysis across diverse carrier frequencies by deep learning

Moiré tomography stands as a potent technique for capturing three-dimensional flow fields, and its precision and accuracy hinge upon the efficiency of moiré fringe analysis. In this paper, a deep learning moiré fringe analysis (DLMFA) method is proposed, and the wrapped phase information could be predicted and analyzed by training moiré fringe datasets across various carrier frequencies. The methodology involves acquiring eight sets of moiré fringes with distinct carrier frequencies through a moiré tomography system. Subsequently, the real and imaginary components of the first-order spectrum of moiré fringes were extracted using Fourier analysis, forming the training datasets for the deep learning model. The trained deep learning model can accurately predict the wrapped phase information corresponding to the moiré fringe’s carrier frequency. The results show that the convergence rate of training loss and validation loss of deep learning model is gradually faster, the prediction loss is gradually reduced, and the structural similarity is also weaker with the increase of carrier frequency. This indicates the impact of carrier frequency on the model’s predictive accuracy robustness against interference, including occlusion and noise. The proposed approach exhibits high accuracy and robustness in moiré fringe analysis, demonstrating applicability across diverse carrier frequencies for flow fields measurement. This study introduces a fresh perspective and solution for the intelligent advancement of moiré tomography.

ShaleSeg: Deep-learning dataset and models for practical fracture segmentation of large-scale shale CT images

Shale gas plays a crucial role as an exploration resource for human life, and its contribution to energy production has been steadily increasing. The success of hydraulic fracturing operations is highly dependent on the understanding of shale fracture propagation at large scales. Currently, the segmentation of large-scale shale fractures is primarily carried out through time-consuming manual or semi-automated methods. The identification of shale fractures is similar to image segmentation, in which deep-learning holds the potential to achieve accurate performance close to human experts. However, its application encounters challenges when dealing with large-scale shale images due to factors such as a lack of datasets, severe artifacts, and limited sample sizes. In this study, we compiled the first publicly available large-scale shale CT dataset, which consists of a total of 4313 scanned images, with over 700 images coming from scanning experiments on 300 cm3 cubic shale samples. We also proposed a new data augmentation method for shale CT images to overcome challenges posed by severe artifacts and small-sample sizes. This method involves simulating image artifacts and reducing high edge gradients along the sample edges, which has proven to be effective in facilitating the efficient training of deep-learning models. We've also conducted several ablation studies to determine the best model design, namely ShaleSeg. ShaleSeg was able to accurately identify multi-scale fractures, achieving a precision of 0.84 and an F-score of 0.81, which is close to the manual results. In the case of minor fractures, precision increased by an increase significantly from 0.191 to 0.513, compared to the traditional algorithm. ShaleSeg significantly improves the accuracy and efficiency of fracture segmentation for large-scale shale images. This advancement will be crucial for a thorough comprehension of the fracture mechanism in shale.

Simulation-driven design of smart gloves for gesture recognition

Smart gloves are in high demand for entertainment, manufacturing, and rehabilitation. However, designing smart gloves has been complex and costly due to trial and error. We propose an open simulation platform for designing smart gloves, including optimal sensor placement and deep learning models for gesture recognition, with reduced costs and manual effort. Our pipeline starts with 3D hand pose extraction from videos and extends to the refinement and conversion of the poses into hand joint angles based on inverse kinematics, the sensor placement optimization based on hand joint analysis, and the training of deep learning models using simulated sensor data. In comparison to the existing platforms that always require precise motion data as input, our platform takes monocular videos, which can be captured with widely available smartphones or web cameras, as input and integrates novel approaches to minimize the impact of the errors induced by imprecise motion extraction from videos. Moreover, our platform enables more efficient sensor placement selection. We demonstrate how the pipeline works and how it delivers a sensible design for smart gloves in a real-life case study. We also evaluate the performance of each building block and its impact on the reliability of the generated design.

Addressing label noise for electronic health records: insights from computer vision for tabular data

The analysis of extensive electronic health records (EHR) datasets often calls for automated solutions, with machine learning (ML) techniques, including deep learning (DL), taking a lead role. One common task involves categorizing EHR data into predefined groups. However, the vulnerability of EHRs to noise and errors stemming from data collection processes, as well as potential human labeling errors, poses a significant risk. This risk is particularly prominent during the training of DL models, where the possibility of overfitting to noisy labels can have serious repercussions in healthcare. Despite the well-documented existence of label noise in EHR data, few studies have tackled this challenge within the EHR domain. Our work addresses this gap by adapting computer vision (CV) algorithms to mitigate the impact of label noise in DL models trained on EHR data. Notably, it remains uncertain whether CV methods, when applied to the EHR domain, will prove effective, given the substantial divergence between the two domains. We present empirical evidence demonstrating that these methods, whether used individually or in combination, can substantially enhance model performance when applied to EHR data, especially in the presence of noisy/incorrect labels. We validate our methods and underscore their practical utility in real-world EHR data, specifically in the context of COVID-19 diagnosis. Our study highlights the effectiveness of CV methods in the EHR domain, making a valuable contribution to the advancement of healthcare analytics and research.

A lightweight and privacy preserved federated learning ecosystem for analyzing verbal communication emotions in identical and non-identical databases

The lack of vocal emotional expression is a major deficit in social communication disorders. The current scenario of artificial intelligence focuses on collaborative training of deep learning models without losing data privacy. The primary objective of this paper is to propose a federated learning-based classification model to identify and analyze the emotional capabilities of individuals with vocal emotion deficits. The methodology has developed a collaborative and privacy-preserved approach using federated learning for training the deep learning models. The proposed methodology utilizes Mel-frequency Cepstral Coefficients (MFCC) to preprocess audio recordings. The four datasets (RAVDESS, CREMA, TESS, SAVEE) including emotion-based classified audio recordings were collected from open sources. The collected audio recordings are 3 s each and the total data set has 668376 audio files with happy - 175119 files, sad – 172611 files, angry – 176346 files, and normal - 144300 files. Further, the input audio was pre-processed to generate MFCC features. The study began with extracting features from multiple pre-trained DL models as its base model. Then, the performance of the federated learning (FL) model was tested on independent and identically distributed (IID) and non-IID data. Further, this paper presents a federated deep learning-based multimodal system for verbal communication emotions classification that uses audio datasets to meet data privacy requirements by DL on the FL ecosystem. As per the findings, the federated learning trained model provides nearly similar parametric results in comparison to base model training. For IID data, the model had 99.71 % validation accuracy, precision (99.73 %), recall (99.69 %), and validation loss (0.01). The FL architecture with non-IID data outperformed these measures with validation accuracy (99.97 %), precision (99.97 %), recall (99.97 %), and least loss (0). Hence the acquired results support the utilization of federated learning ecosystem-based trained models with identically and non-identically distributed audio features from emotion identification without losing parametric results. In conclusion, the proposed techniques could be applied to identify verbal emotional deficits in individuals and could support developing emerging technological interventions for their well-being.

Mitigating human fall injuries: A novel system utilizing 3D 4-stream convolutional neural networks and image fusion

Unintentional human falls, especially in seniors, lead to serious injuries, fatalities, and reduced standard of life. Vision-based fall detection methods have demonstrated their usefulness in timely fall response, helping to lessen such injuries. This paper presents an automated vision-based fall detection system that triggers immediate fall reporting. By incorporating human segmentation and image fusion in the pre-processing stage, the system enhances the accuracy of human action classification, thereby ensuring precise fall alerts. It further employs the innovative 4-stream 3D convolutional neural network (4S-3DCNN) model to learn different but consecutive spatial and temporal features. The system processes video input or live surveillance, segmenting human presence every 32 frames using a fine-tuned deep-learning model and applying a three-level image fusion to accentuate movement differences. This technique produces four pre-processed images, input to the 4S-3DCNN model for classification. Consecutive detection of “Falling” and “Fallen” actions triggers an alert for immediate intervention. The original 4S-3DCNN model is an end-to-end trained deep learning model with a fully connected layer serving as a classifier. The research also evaluates the performance of combining the 4S-3DCNN model with Autoencoders and Support Vector Machines (SVM) networks as classifiers. The SVM classifier demonstrated ideal fall detection performance with 100% accuracy using the MCFD, URFD, and Le2i FDD datasets. The proposed system is vital for detecting and preventing falls and reducing healthcare expenses and productivity losses.