{"id":{"repo_id":"uic","oai_identifier":"oai:figshare.com:article/32991815"},"canonical_url":"https://search.dev.ndltd.org/etd/uic/oai:figshare.com:article/32991815","repository":{"repo_id":"uic","name":"University of Illinois - Chicago","base_url":"https://api.figshare.com/v2/oai"},"display":{"title":"Trainable Pre-Filtering for Deep Neural Networks and Applications","abstract":"Deep learning has achieved strong performance in vision-based perception and prediction tasks; however, many modern deep neural networks depend heavily on large labeled datasets and purely data-driven feature learning, which limits robustness, efficiency, and real-world deployability, especially in noisy, data-scarce, and resource-constrained environments. This dissertation proposes a unified framework that integrates trainable pre-filtering layers inspired by classical signal processing directly into deep neural network architectures. The framework embeds physically meaningful operations—such as directional filtering, Sobel gradients, Gaussian and low-pass smoothing, multi-resolution representations, and orthogonal transforms—into the learning pipeline, introducing explicit inductive bias while preserving end-to-end optimization. The proposed approach is validated across multiple real-world applications, including infrared air-leak identification, adversarially robust image classification, power line monitoring, and next-day wildfire spread prediction. Experimental results demonstrate consistent performance improvements, enhanced robustness, and reduced model complexity compared to conventional deep learning baselines. These results show that domain-informed, trainable pre-filtering can effectively improve convergence, generalization, robustness, and computational efficiency, particularly under limited data availability, noise, and adversarial perturbations. Unlike fixed preprocessing techniques, the proposed filters are jointly optimized with network parameters, enabling task-adaptive behavior while maintaining clear physical interpretability. Overall, this work establishes trainable pre-filtering as a principled framework for bridging classical signal processing and modern deep learning in environmental hazard analysis and infrastructure monitoring.","abstract_html":"Deep learning has achieved strong performance in vision-based perception and prediction tasks; however, many modern deep neural networks depend heavily on large labeled datasets and purely data-driven feature learning, which limits robustness, efficiency, and real-world deployability, especially in noisy, data-scarce, and resource-constrained environments. This dissertation proposes a unified framework that integrates trainable pre-filtering layers inspired by classical signal processing directly into deep neural network architectures. The framework embeds physically meaningful operations—such as directional filtering, Sobel gradients, Gaussian and low-pass smoothing, multi-resolution representations, and orthogonal transforms—into the learning pipeline, introducing explicit inductive bias while preserving end-to-end optimization. The proposed approach is validated across multiple real-world applications, including infrared air-leak identification, adversarially robust image classification, power line monitoring, and next-day wildfire spread prediction. Experimental results demonstrate consistent performance improvements, enhanced robustness, and reduced model complexity compared to conventional deep learning baselines. These results show that domain-informed, trainable pre-filtering can effectively improve convergence, generalization, robustness, and computational efficiency, particularly under limited data availability, noise, and adversarial perturbations. Unlike fixed preprocessing techniques, the proposed filters are jointly optimized with network parameters, enabling task-adaptive behavior while maintaining clear physical interpretability. Overall, this work establishes trainable pre-filtering as a principled framework for bridging classical signal processing and modern deep learning in environmental hazard analysis and infrastructure monitoring.","abstract_has_math":false,"creators":["Shuaiang Rong (11408235)"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-07-18T03:24:25Z","date_published":"2026-07-18T03:24:25Z","updated_at":"2026-07-27T21:33:04Z","subjects":["Electrical and Computer Engineering","Artificial Intelligence","Signal Processing","Computer Vision"],"languages":[],"rights":["In Copyright"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://doi.org/10.25417/uic.32991815.v2","outbound_label":"DOI","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:creator","label":"Author","values":["Shuaiang Rong (11408235)"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2026-07-18T03:24:25Z"]},{"key":"dc:relation","label":"Dc Relation","values":["https://figshare.com/articles/thesis/Trainable_Pre-Filtering_for_Deep_Neural_Networks_and_Applications/32991815"]},{"key":"dc:type","label":"Dc Type","values":["Text","Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Electrical and Computer Engineering","Artificial Intelligence","Signal Processing","Computer Vision"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["In Copyright"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["10.25417/uic.32991815.v2"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Deep learning has achieved strong performance in vision-based perception and prediction tasks; however, many modern deep neural networks depend heavily on large labeled datasets and purely data-driven feature learning, which limits robustness, efficiency, and real-world deployability, especially in noisy, data-scarce, and resource-constrained environments. This dissertation proposes a unified framework that integrates trainable pre-filtering layers inspired by classical signal processing directly into deep neural network architectures. The framework embeds physically meaningful operations—such as directional filtering, Sobel gradients, Gaussian and low-pass smoothing, multi-resolution representations, and orthogonal transforms—into the learning pipeline, introducing explicit inductive bias while preserving end-to-end optimization. The proposed approach is validated across multiple real-world applications, including infrared air-leak identification, adversarially robust image classification, power line monitoring, and next-day wildfire spread prediction. Experimental results demonstrate consistent performance improvements, enhanced robustness, and reduced model complexity compared to conventional deep learning baselines. These results show that domain-informed, trainable pre-filtering can effectively improve convergence, generalization, robustness, and computational efficiency, particularly under limited data availability, noise, and adversarial perturbations. Unlike fixed preprocessing techniques, the proposed filters are jointly optimized with network parameters, enabling task-adaptive behavior while maintaining clear physical interpretability. Overall, this work establishes trainable pre-filtering as a principled framework for bridging classical signal processing and modern deep learning in environmental hazard analysis and infrastructure monitoring."]},{"key":"dc:title","label":"Title","values":["Trainable Pre-Filtering for Deep Neural Networks and Applications"]}]}],"canonical_facts":{"dc:creator":["Shuaiang Rong (11408235)"],"dc:date":["2026-07-18T03:24:25Z"],"dc:description":["Deep learning has achieved strong performance in vision-based perception and prediction tasks; however, many modern deep neural networks depend heavily on large labeled datasets and purely data-driven feature learning, which limits robustness, efficiency, and real-world deployability, especially in noisy, data-scarce, and resource-constrained environments. This dissertation proposes a unified framework that integrates trainable pre-filtering layers inspired by classical signal processing directly into deep neural network architectures. The framework embeds physically meaningful operations—such as directional filtering, Sobel gradients, Gaussian and low-pass smoothing, multi-resolution representations, and orthogonal transforms—into the learning pipeline, introducing explicit inductive bias while preserving end-to-end optimization. The proposed approach is validated across multiple real-world applications, including infrared air-leak identification, adversarially robust image classification, power line monitoring, and next-day wildfire spread prediction. Experimental results demonstrate consistent performance improvements, enhanced robustness, and reduced model complexity compared to conventional deep learning baselines. These results show that domain-informed, trainable pre-filtering can effectively improve convergence, generalization, robustness, and computational efficiency, particularly under limited data availability, noise, and adversarial perturbations. Unlike fixed preprocessing techniques, the proposed filters are jointly optimized with network parameters, enabling task-adaptive behavior while maintaining clear physical interpretability. Overall, this work establishes trainable pre-filtering as a principled framework for bridging classical signal processing and modern deep learning in environmental hazard analysis and infrastructure monitoring."],"dc:identifier":["10.25417/uic.32991815.v2"],"dc:relation":["https://figshare.com/articles/thesis/Trainable_Pre-Filtering_for_Deep_Neural_Networks_and_Applications/32991815"],"dc:rights":["In Copyright"],"dc:subject":["Electrical and Computer Engineering","Artificial Intelligence","Signal Processing","Computer Vision"],"dc:title":["Trainable Pre-Filtering for Deep Neural Networks and Applications"],"dc:type":["Text","Thesis"]},"updated_at":"2026-07-27T21:33:04Z"}