Publications

A novel Auto-Regressive image generation approach that models images as hierarchical compositions of interpretable visual layers. CART decomposes image generation into structured, auto-regressive predictions, enabling more controllable and interpretable generation.

Introduces Volterra Neural Ordinary Differential Equations — a piecewise continuous neural network architecture for image classification that extends the Volterra series formulation into the continuous-time domain.

Presents GalaxyEdit, a large-scale image editing dataset paired with an enhanced diffusion adapter for controllable, high-fidelity image editing using generative diffusion models.

Proposes DCT-Shield, a frequency-domain defense that protects images against malicious AI-based editing by embedding robust perturbations in the DCT domain, preventing unauthorized manipulation while preserving visual quality.

LLVD leverages LSTM-based explicit motion modeling in a compressed latent space for blind video denoising, enabling effective temporal coherence without prior knowledge of noise statistics.

Vid-Freeze protects images from malicious image-to-video generation by introducing temporal freezing perturbations that disrupt video synthesis models while remaining imperceptible to human viewers.

A restoration network based on the Volterra series formulation, incorporating non-linearity through higher-order convolutions instead of activation functions. Achieves state-of-the-art performance in Image/Video Restoration and establishes NAF-NET as a special case of the Volterra Neural Network class.

A Volterra filter-inspired network architecture introducing controlled non-linearities through interactions between delayed input samples. The cascaded parallel implementation reduces parameters significantly while outperforming state-of-the-art CNN approaches on UCF-101 and HMDB-51 for action recognition.

Addresses unsupervised Domain Adaptation using optimal transport theory with a verifiably efficient solution to learn the best latent feature representation, by minimizing the cost of transporting target domain samples to the source domain distribution.

A deep structure encoder using VNNs to seek a latent representation of multi-modal data from a union of subspaces. Shows significant improvement in clustering performance over CNN auto-encoders, with improved sample complexity and robust classification.

A multi-scale no-reference IQA method that is aware of different degradation types, enabling better perceptual quality estimation without requiring a pristine reference image.

A Volterra filter-inspired network that reduces CNN complexity via controlled non-linearities from delayed input interactions. The cascaded parallel implementation significantly reduces parameters while outperforming state-of-the-art CNN approaches on UCF-101 and HMDB-51.

A conditional GAN-based multi-modal sensor fusion framework where generative models are trained to commute, enabling consistent cross-modal synthesis and robust fusion even under sensor degradation.

A data-driven approach to multi-modal fusion exploiting optimal features from a latent space learned by a generative network conditioned on individual modalities, used to detect damaged sensors and safeguard fused system performance.

Exploits events observed by different sensors to detect and classify objects by exploring feature dependence across modalities and generating informed probability distributions. Also addresses damaged sensors by learning a hidden inter-modal space.

Applies machine learning methods to predict Atlantic hurricane activity, demonstrating that data-driven models can capture the complex atmospheric dynamics governing hurricane formation and intensity.

Uses a Conditional GAN to distill knowledge across sensor modalities, generating representative information from missing modalities given available ones. Achieves better performance than traditional approaches and teacher-student models.

Combines sensor-observed events to detect and classify objects by exploring feature dependence across sensors and generating more informed probability distributions over events.

A multi-scale approach to detecting skin pixels in images, with applications to video privacy protection and content moderation systems.