Personalized fall detection models can significantly improve accuracy by adapting to individual motion patterns, yet their effectiveness is often limited by the scarcity of real-world fall data and the dominance of non-fall feedback samples. This imbalance biases the model toward routine activities and weakens its sensitivity to true fall events. To address this challenge, we propose a personalization framework that combines semi-supervised clustering with contrastive learning to identify and balance the most informative user feedback samples. The framework is evaluated under three retraining strategies, including Training from Scratch (TFS), Transfer Learning (TL), and Few-Shot Learning (FSL), to assess adaptability across learning paradigms. Real-time experiments with ten participants show that the TFS approach achieves the highest performance, with up to a 25% improvement over the baseline, while FSL achieves the second-highest performance with a 7% improvement, demonstrating the effectiveness of selective personalization for real-world deployment.

Time-series data is widely used across domains but presents challenges due to high dimensionality and noise. This study proposes a hybrid feature selection approach to enhance forecasting accuracy by reducing irrelevant features. We first trained a 1D-CNN-based forecasting model using all features as a baseline. Then, we evaluated four feature selection methods: two traditional (Variance and Dynamic Mode Decomposition) and two XAI-based (SHAP and LIME). To address the limitations of individual methods, we introduced a hybrid approach combining Variance and LIME. Experiments on four diverse datasets, including EEG (seizure forecasting), Pole-balancing (pre-fall detection), Weather (meteorological forecasting), and Electricity (demand prediction), show that the hybrid method consistently outperformed all others. It achieved the highest gains in EEG (+7.34%), Pole-balancing (+9.02%), Weather (+8.62%), and Electricity (+6.72%), demonstrating robust, interpretable, and generalizable performance across tasks.

The lack of real-world data in clinical fields poses a major obstacle in training effective AI models for diagnostic and preventive tools in medicine. Generative AI has shown promise in increasing data volume and enhancing model training, particularly in computer vision and natural language processing (NLP) domains. However, generating physiological time-series data, a common type in medical AI applications, presents unique challenges due to its inherent complexity and variability. This paper introduces TransConv-DDPM, an enhanced generative AI method for biomechanical and physiological time-series data generation. The model employs a denoising diffusion probabilistic model (DDPM) with U-Net, multi-scale convolution modules, and a transformer layer to capture both global and local temporal dependencies. We evaluated TransConv-DDPM on three diverse datasets, generating both long and short-sequence time-series data. Quantitative comparisons against state-of-the-art methods, TimeGAN and Diffusion-TS, using four performance metrics, demonstrated promising results, particularly on the SmartFallMM and EEG datasets, where it effectively captured the more gradual temporal change patterns between data points. Additionally, a utility test on the SmartFallMM dataset revealed that adding synthetic fall data generated by TransConv-DDPM improved predictive model performance, showing a 13.64% improvement in F1-score and a 14.93% increase in overall accuracy compared to the baseline model trained solely on fall data from the SmartFallMM dataset. These findings highlight the potential of TransConv-DDPM to generate high-quality synthetic data for real-world applications.

Light Fields (LF) capture both angular and spatial information of light rays traveling through free space, resulting in a richer representation of a scene from multiple viewpoints. However, the high dimensionality of LF data poses a challenge for compression and transmission algorithms, which can degrade visual quality. To address this issue, we propose a novel no-reference LF Image Quality Assessment (LF-IQA) method that accurately predicts the quality of complex LF content. Our method is based on a diverse neural network architecture that includes Convolutional Neural Network (CNN) blocks, Atrous Convolutional Layers (ACLs), and Long Short-Term Memory (LSTM) layers. The model architecture consists of two streams, each containing CNN, ACL, and LSTM layers, which take the horizontal and vertical epipolar plane LF images as input. CNN blocks extract basic input features, while ACL blocks extract high-level features. The LSTM layers are used to capture long-term dependencies and relationships among distortion-related features. Finally, the outputs of both streams are concatenated and fed into the regression block for quality prediction. Our results demonstrate that the proposed LF-IQA method is robust and outperforms current state-of-the-art methods, even for complex LF content.

Unlike regular images that represent only light intensities, Light Field (LF) contents carry information about the intensity of light in a scene, including the direction in which light rays are traveling in space. This allows for a richer representation of our world, but requires large amounts of data that need to be processed and compressed before being transmitted to the viewer. Since these techniques may introduce distortions, the design of Light Field Image Quality Assessment (LF-IQA) methods is important. Currently, most LF-IQA methods use traditional 2D image quality assessment techniques or rely on low-level spatial features. In this paper, we propose a novel no-reference LF-IQA method that takes into account both LF angular and spatial information. The proposed method is made up of two processing streams with identical blocks of Convolutional Neural Network (CNN), Atrous Convolution layers (ACL), and a regression block for quality prediction. The results show that the method is robust and outperforms current state-of-the-art methods.

This paper contains the research proposal of Sana Alamgeer that was presented at the MMSys 2022 doctoral symposium. Unlike regular images that represent only light intensities, Light Field (LF) contents carry information about the intensity of light in a scene, including the direction light rays are traveling in space. This allows for a richer representation of our world, but requires large amounts of data that need to be processed and compressed before being transmitted to the viewer. Since these techniques may introduce distortions, the design of Light Field Image Quality Assessment (LF-IQA) methods is important. The majority of LF-IQA methods based on traditional Convolutional Neural Network (CNN) have limitations, i.e. they are unable to increase the receptive field of a neuron-pixel to model non-local image features. In this work, we propose a novel no-reference LF-IQA method that is based on Deep Graph Convolutional Neural Network (GCNN). Our method not only takes into account both LF angular and spatial information, but also learns the order of pixel information. Specifically, the method is composed of one input layer that takes a pair of graphs and their corresponding subjective quality scores as labels, 4 GCNN layers, fully connected layers, and a regression block for quality prediction. Our aim is to develop the quality prediction method with maximum accuracy for distorted LF content.

Light Field (LF) cameras capture angular and spa-tial information and, consequently, require a large amount of resources in memory and bandwidth. To reduce these requirements, LF contents generally need to undergo compression and transmission protocols. Since these techniques may introduce distortions, the design of Light-Field Image Quality Assessment (LFI - IQA) methods are important to monitor the quality of the LF image (LFI) content at the user side. The majority of the existing LFI-IQA methods work in the spatial domain, where it is more difficult to analyze changes in the spatial and angular domains. In this work, we present a novel NR LFI-IQA, which is based on a Deep Neural Network that uses Frequency domain inputs (DNNF-LFIQA). The proposed method predicts the quality of an LF image by taking as input the Fourier magnitude spectrum of LF contents, represented as horizontal and ver-tical Epipolar Plane Images (EPI)s. Specifically, DNNF-LFIQA is composed of two processing streams (streaml and stream2) that take as inputs the horizontal and the vertical epipolar plane images in the frequency domain. Both streams are composed of identical blocks of convolutional neural networks (CNNs), with their outputs being combined using two fusion blocks. Finally, the fused feature vector is fed to a regression block to generate the quality prediction. Results show that the proposed method is fast, robust, and accurate.

Due to the rapid development of multimedia technologies, over the last decades image quality assessment (IQA) has become an important topic. As a consequence, a great research effort has been made to develop computational models that estimate image quality. Among the possible IQA approaches, blind IQA (BIQA) is of fundamental interest as it can be used in most multimedia applications. BIQA techniques measure the perceptual quality of an image without using the reference (or pristine) image. This paper proposes a new BIQA method that uses a combination of texture features and saliency maps of an image. Texture features are extracted from the images using the local binary pattern (LBP) operator at multiple scales. To extract the salient of an image, i.e. the areas of the image that are the main attractors of the viewers' attention, we use computational visual attention models that output saliency maps. These saliency maps can be used as weighting functions for the LBP maps at multiple scales. We propose an operator that produces a combination of multiscale LBP maps and saliency maps, which is called the multiscale salient local binary pattern (MSLBP) operator. To define which is the best model to be used in the proposed operator, we investigate the performance of several saliency models. Experimental results demonstrate that the proposed method is able to estimate the quality of impaired images with a wide variety of distortions. The proposed metric has a better prediction accuracy than state-of-the-art IQA methods.