@prefix ocrer: . @prefix owl: . @prefix scires: . @prefix xsd: . @prefix skos: . @prefix rdfs: . @prefix ocresd: . @prefix swo: . @prefix cito: . @prefix geo: . @prefix ocresst: . @prefix dcterms: . @prefix vivo: . @prefix event: . @prefix vann: . @prefix foaf: . @prefix c4o: . @prefix fabio: . @prefix vcard: . @prefix thkoeln: . @prefix vitro: . @prefix vitro-public: . @prefix rdf: . @prefix ocresp: . @prefix bibo: . @prefix obo: . @prefix ro: . obo:BFO_0000031 a owl:Class ; rdfs:label "Bedingter Kontinuant"@de-DE , "Generically Dependent Continuant"@en-US , "Generically Dependent Continuant"@en . . a thkoeln:Zitat , owl:Thing . a , owl:NamedIndividual , owl:Thing ; rdfs:label "Open Access"@en-US , "Open Access"@de-DE . obo:BFO_0000002 a owl:Class ; rdfs:label "Kontinuant"@de-DE , "Continuant"@en-US , "Continuant"@en . a thkoeln:Zitat , owl:Thing . obo:BFO_0000001 a owl:Class ; rdfs:label "Entität"@de-DE , "Entity"@en-US , "Entity"@en . a owl:NamedIndividual , , owl:Thing ; rdfs:label "Text"@de-DE , "Text"@en-US . bibo:Document a owl:Class ; rdfs:label "Dokument"@de-DE , "Document"@en-US , "Document"@en . thkoeln:Konferenzpaper a owl:Class ; rdfs:label "Conference Paper"@en-US , "Konferenzpaper"@de-DE . owl:Thing a owl:Class . a thkoeln:Organisationseinheit , foaf:Agent , obo:BFO_0000001 , obo:BFO_0000004 , obo:BFO_0000002 , foaf:Organization , owl:Thing ; rdfs:label "Institut für Medien- und Phototechnik" ; . obo:IAO_0000030 a owl:Class ; rdfs:label "Entität für Informationsinhalt"@de-DE , "Information Content Entity"@en-US , "Information Content Entity"@en . a foaf:Document , obo:BFO_0000002 , obo:BFO_0000031 , obo:IAO_0000030 , bibo:Document , thkoeln:Konferenzpaper , owl:Thing , obo:BFO_0000001 , thkoeln:Publikation ; rdfs:label "RDF description of DEEP-SEED: From Scratch to Ensemble -- A Deep Learning Approach to Seedling Classification - https://fis.th-koeln.de/vivo/individual/publ_20406" , "DEEP-SEED: From Scratch to Ensemble -- A Deep Learning Approach to Seedling Classification" ; thkoeln:abstractDE "This paper evaluates and compares three deep learning (DL) approaches for plant seedlings classification using a dataset consisting of 4750 images of 12 different plant species. Several DL approaches were considered, including a custom convolutional neural network (CNN) trained from scratch, a pre-trained CNN \"ResNet-18\" and a pre-trained vision transformer (ViT) \"vit-base-patch16-224\" fine-tuned for the task at hand. To address the challenges of data scarcity and class imbalance, extensive data augmentation techniques such as random rotations, flips and color jittering were employed. Results showed that transfer learning with ResNet-18 outperforms the custom model, achieving a mean F1-score (micro-averaged) of ~0.961 on the test set. The custom CNN, still achieved a competitive F1-score of ~0.927, demonstrating that even smaller locally trained architectures can be viable if carefully designed and thoroughly regularized. While the ViT model achieved the highest F1-score of ~0.967, an ensemble combining the predictions of all three models outperformed the single models with a score of ~0.971. Finally, potential improvements are outlined, including deeper architectures, synthetic image generation and interpretability measures, to further improve seedling classification performance." ; thkoeln:sprache "englisch" ; thkoeln:status "validiert" ; ; ; ; ; , , ; ; "2026-04-04T01:48:41"^^xsd:dateTime ; ; ; bibo:abstract "This paper evaluates and compares three deep learning (DL) approaches for plant seedlings classification using a dataset consisting of 4750 images of 12 different plant species. Several DL approaches were considered, including a custom convolutional neural network (CNN) trained from scratch, a pre-trained CNN \"ResNet-18\" and a pre-trained vision transformer (ViT) \"vit-base-patch16-224\" fine-tuned for the task at hand. To address the challenges of data scarcity and class imbalance, extensive data augmentation techniques such as random rotations, flips and color jittering were employed. Results showed that transfer learning with ResNet-18 outperforms the custom model, achieving a mean F1-score (micro-averaged) of ~0.961 on the test set. The custom CNN, still achieved a competitive F1-score of ~0.927, demonstrating that even smaller locally trained architectures can be viable if carefully designed and thoroughly regularized. While the ViT model achieved the highest F1-score of ~0.967, an ensemble combining the predictions of all three models outperformed the single models with a score of ~0.971. Finally, potential improvements are outlined, including deeper architectures, synthetic image generation and interpretability measures, to further improve seedling classification performance." ; bibo:doi "10.57684/COS-1297" ; bibo:isbn13 "#MODS-ID.bibthk_mods_00014907" , "#defaultkonferenzband" ; bibo:pageEnd "28" ; bibo:pageStart "18" ; vitro:mostSpecificType thkoeln:Konferenzpaper ; vivo:dateTimeValue ; vivo:freetextKeyword "Convolutional neural networks" , "Vision transformers" , "Machine learning" , "Image classification" , "Transfer learning" ; vivo:relatedBy . . a obo:BFO_0000003 , obo:BFO_0000008 , obo:BFO_0000001 , vivo:DateTimeValue , obo:BFO_0000148 , owl:Thing . a obo:BFO_0000001 , obo:BFO_0000020 , vivo:Authorship , vivo:Relationship , obo:BFO_0000002 , owl:Thing ; vivo:relates . thkoeln:Publikation a owl:Class ; rdfs:label "Publikationen"@de-DE , "Publications"@en-US . a thkoeln:Zitat , owl:Thing .