{"id":{"repo_id":"sask","oai_identifier":"oai:harvest.usask.ca:10388/17303"},"canonical_url":"https://search.dev.ndltd.org/etd/sask/oai:harvest.usask.ca:10388/17303","repository":{"repo_id":"sask","name":"University of Saskatchewan","base_url":"https://harvest.usask.ca/server/oai/request"},"display":{"title":"RGB and NIR Image Registration for Proximal Multiband Imaging","abstract":"Accurate monitoring of plant traits is critical to advancing agricultural productivity and sustainability. Multispectral imaging relying on red-green-blue (RGB) and near-infrared (NIR) imagery has become an essential tool in high-throughput plant phenotyping. However, accurate analysis requires precise alignment between the images captured using these different modalities. This alignment process is known as image registration, a task complicated by geometric distortions, radiometric differences, and occlusions derived from the sensors' placement. These challenges are especially pronounced in close-to-ground proximal sensing platforms where the RGB and NIR sensors are separated by a baseline distance. This thesis addresses these challenges by first identifying a set of similarity metrics capable of quantifying spatial misalignment in RGB/NIR image pairs. These metrics were assessed under a set of evaluation criteria using a synthetically generated dataset that modeled increasing baseline distances. The Census transform, Gabor filter, and Color Index of Vegetation Extraction (CIVE) vegetation index were used as preprocessing filters to minimize radiometric differences and enhance cross-spectral comparability. It was found that metrics using the Gabor and CIVE filters exhibited increased metric performance, with CIVE negatively affected by the absence of green vegetation. The best-performing metrics include RMSE, ERGAS, LPIPS, VIF, and SSIM, for which their scores under perfect NIR/RGB spatial similarity were successfully modeled. A two-step registration pipeline consisting of global alignment and dense optical flow-based local warping was developed and benchmarked across synthetic, controlled, and field datasets. It was found that using a phase cross-correlation approach was more consistent than traditional feature-based algorithms, and that applying the mentioned filters as a preprocessing step affected the performance of the optical flow algorithm. Additionally, a modified approach for occlusion detection via forward-backward consistency checks was implemented, showing that occlusions can be successfully identified using dense optical flow even in closely compacted scenes. Overall, the work demonstrates that through the integration of preprocessing, robust similarity quantification, and occlusion-aware dense registration, accurate alignment of RGB/NIR imagery is achievable and quantifiable, even in the absence of ground truth.","abstract_html":"Accurate monitoring of plant traits is critical to advancing agricultural productivity and sustainability. Multispectral imaging relying on red-green-blue (RGB) and near-infrared (NIR) imagery has become an essential tool in high-throughput plant phenotyping. However, accurate analysis requires precise alignment between the images captured using these different modalities. This alignment process is known as image registration, a task complicated by geometric distortions, radiometric differences, and occlusions derived from the sensors' placement. These challenges are especially pronounced in close-to-ground proximal sensing platforms where the RGB and NIR sensors are separated by a baseline distance. This thesis addresses these challenges by first identifying a set of similarity metrics capable of quantifying spatial misalignment in RGB/NIR image pairs. These metrics were assessed under a set of evaluation criteria using a synthetically generated dataset that modeled increasing baseline distances. The Census transform, Gabor filter, and Color Index of Vegetation Extraction (CIVE) vegetation index were used as preprocessing filters to minimize radiometric differences and enhance cross-spectral comparability. It was found that metrics using the Gabor and CIVE filters exhibited increased metric performance, with CIVE negatively affected by the absence of green vegetation. The best-performing metrics include RMSE, ERGAS, LPIPS, VIF, and SSIM, for which their scores under perfect NIR/RGB spatial similarity were successfully modeled. A two-step registration pipeline consisting of global alignment and dense optical flow-based local warping was developed and benchmarked across synthetic, controlled, and field datasets. It was found that using a phase cross-correlation approach was more consistent than traditional feature-based algorithms, and that applying the mentioned filters as a preprocessing step affected the performance of the optical flow algorithm. Additionally, a modified approach for occlusion detection via forward-backward consistency checks was implemented, showing that occlusions can be successfully identified using dense optical flow even in closely compacted scenes. Overall, the work demonstrates that through the integration of preprocessing, robust similarity quantification, and occlusion-aware dense registration, accurate alignment of RGB/NIR imagery is achievable and quantifiable, even in the absence of ground truth.","abstract_has_math":false,"creators":["Vizcaino Ramirez, David Fernando"],"institution":"University of Saskatchewan","degree_name":"Master of Science (M.Sc.)","degree_level":"Masters","degree_discipline":"Mechanical Engineering","degree_department":null,"school":null,"contributors":[],"advisors":["Noble, Scott"],"committee_chairs":[],"committee_members":["McWalter, Emily","Berscheid, Brian"],"year":2025,"date_issued":"2025-09-17","date_published":"2025-09-17","updated_at":"2026-07-24T04:26:54Z","subjects":["RGB","NIR","Image registration","multiband","Proximal sensing","Optical flow","Census","Gabor","CIVE"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/10388/17303","outbound_label":"Handle","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Noble, Scott"]},{"key":"dc:contributor.committeemember","label":"Committee Member","values":["McWalter, Emily","Berscheid, Brian"]},{"key":"dc:creator","label":"Author","values":["Vizcaino Ramirez, David Fernando"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2025-09-17T14:58:34Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2025-09-17T14:58:34Z"]},{"key":"dc:date.issued","label":"Date","values":["2025-09-17"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Mechanical Engineering"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Masters"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master of Science (M.Sc.)"]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Saskatchewan"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["RGB","NIR","Image registration","multiband","Proximal sensing","Optical flow","Census","Gabor","CIVE"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/10388/17303"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Accurate monitoring of plant traits is critical to advancing agricultural productivity and sustainability. Multispectral imaging relying on red-green-blue (RGB) and near-infrared (NIR) imagery has become an essential tool in high-throughput plant phenotyping. However, accurate analysis requires precise alignment between the images captured using these different modalities. This alignment process is known as image registration, a task complicated by geometric distortions, radiometric differences, and occlusions derived from the sensors' placement. These challenges are especially pronounced in close-to-ground proximal sensing platforms where the RGB and NIR sensors are separated by a baseline distance. This thesis addresses these challenges by first identifying a set of similarity metrics capable of quantifying spatial misalignment in RGB/NIR image pairs. These metrics were assessed under a set of evaluation criteria using a synthetically generated dataset that modeled increasing baseline distances. The Census transform, Gabor filter, and Color Index of Vegetation Extraction (CIVE) vegetation index were used as preprocessing filters to minimize radiometric differences and enhance cross-spectral comparability. It was found that metrics using the Gabor and CIVE filters exhibited increased metric performance, with CIVE negatively affected by the absence of green vegetation. The best-performing metrics include RMSE, ERGAS, LPIPS, VIF, and SSIM, for which their scores under perfect NIR/RGB spatial similarity were successfully modeled. A two-step registration pipeline consisting of global alignment and dense optical flow-based local warping was developed and benchmarked across synthetic, controlled, and field datasets. It was found that using a phase cross-correlation approach was more consistent than traditional feature-based algorithms, and that applying the mentioned filters as a preprocessing step affected the performance of the optical flow algorithm. Additionally, a modified approach for occlusion detection via forward-backward consistency checks was implemented, showing that occlusions can be successfully identified using dense optical flow even in closely compacted scenes. Overall, the work demonstrates that through the integration of preprocessing, robust similarity quantification, and occlusion-aware dense registration, accurate alignment of RGB/NIR imagery is achievable and quantifiable, even in the absence of ground truth."]},{"key":"dc:format.mimetype","label":"Dc Format Mimetype","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["RGB and NIR Image Registration for Proximal Multiband Imaging"]}]}],"canonical_facts":{"dc:contributor.advisor":["Noble, Scott"],"dc:contributor.committeemember":["McWalter, Emily","Berscheid, Brian"],"dc:creator":["Vizcaino Ramirez, David Fernando"],"dc:date.accessioned":["2025-09-17T14:58:34Z"],"dc:date.available":["2025-09-17T14:58:34Z"],"dc:date.issued":["2025-09-17"],"dc:description.abstract":["Accurate monitoring of plant traits is critical to advancing agricultural productivity and sustainability. Multispectral imaging relying on red-green-blue (RGB) and near-infrared (NIR) imagery has become an essential tool in high-throughput plant phenotyping. However, accurate analysis requires precise alignment between the images captured using these different modalities. This alignment process is known as image registration, a task complicated by geometric distortions, radiometric differences, and occlusions derived from the sensors' placement. These challenges are especially pronounced in close-to-ground proximal sensing platforms where the RGB and NIR sensors are separated by a baseline distance. This thesis addresses these challenges by first identifying a set of similarity metrics capable of quantifying spatial misalignment in RGB/NIR image pairs. These metrics were assessed under a set of evaluation criteria using a synthetically generated dataset that modeled increasing baseline distances. The Census transform, Gabor filter, and Color Index of Vegetation Extraction (CIVE) vegetation index were used as preprocessing filters to minimize radiometric differences and enhance cross-spectral comparability. It was found that metrics using the Gabor and CIVE filters exhibited increased metric performance, with CIVE negatively affected by the absence of green vegetation. The best-performing metrics include RMSE, ERGAS, LPIPS, VIF, and SSIM, for which their scores under perfect NIR/RGB spatial similarity were successfully modeled. A two-step registration pipeline consisting of global alignment and dense optical flow-based local warping was developed and benchmarked across synthetic, controlled, and field datasets. It was found that using a phase cross-correlation approach was more consistent than traditional feature-based algorithms, and that applying the mentioned filters as a preprocessing step affected the performance of the optical flow algorithm. Additionally, a modified approach for occlusion detection via forward-backward consistency checks was implemented, showing that occlusions can be successfully identified using dense optical flow even in closely compacted scenes. Overall, the work demonstrates that through the integration of preprocessing, robust similarity quantification, and occlusion-aware dense registration, accurate alignment of RGB/NIR imagery is achievable and quantifiable, even in the absence of ground truth."],"dc:format.mimetype":["application/pdf"],"dc:identifier.uri":["https://hdl.handle.net/10388/17303"],"dc:language.iso":["en"],"dc:subject":["RGB","NIR","Image registration","multiband","Proximal sensing","Optical flow","Census","Gabor","CIVE"],"dc:title":["RGB and NIR Image Registration for Proximal Multiband Imaging"],"dc:type":["Thesis"],"thesis:degree_discipline":["Mechanical Engineering"],"thesis:degree_level":["Masters"],"thesis:degree_name":["Master of Science (M.Sc.)"],"thesis:institution_name":["University of Saskatchewan"]},"updated_at":"2026-07-24T04:26:54Z"}