Splitter: Faster Inference Through Channel Partitioning and Feature Fusion

Basit Öğe Kaydı
dc.contributor.authorKoyun, Onur Can
dc.contributor.authorEroglu, Kemal Ilgar
dc.contributor.authorTöreyin, Behçet Ugur
dc.date.accessioned2026-04-04T18:48:34Z
dc.date.available2026-04-04T18:48:34Z
dc.date.issued2025
dc.description32nd IEEE International Conference on Image Processing, ICIP 2025 -- 14 September 2025 through 17 September 2025 -- Anchorage -- 216813
dc.description.abstract—This paper presents Splitter, a novel architecture designed to enhance feature extraction and optimize computational efficiency in deep learning models. Splitter employs a unique channel-splitting mechanism that divides input channels into three parallel path; Identity, Activation, and Spatial Mixing to perform distinct operations. By selectively applying spatial mixing via max-pooling or multi-head attention, Splitter balances computational frugality with representational richness. On the ImageNet-1k benchmark, Splitter-S achieves 74.4 % Top-1 accuracy at 9,347 images/s, while Splitter-M and Splitter-L deliver 76.2 % and 78.3 % Top-1 accuracy at 5,893 images/s and 4,719 images/s, respectively. When integrated into a RetinaNet detector on COCO, Splitter-S attains 32.1 % AP (52.4 % AP50, 33.7 % AP75). These results confirm that Splitter matches or surpasses state-of-the-art efficient models while significantly boosting throughput, making it exceptionally well-suited for deployment in resource-limited environments. ©2025 IEEE.
dc.description.sponsorshipTürkiye Bilimsel ve Teknolojik Araştırma Kurumu, TUBITAK, (5239903); Türkiye Bilimsel ve Teknolojik Araştırma Kurumu, TUBITAK; Istanbul Teknik Üniversitesi, IT, (PMA-2024-45912, ITU-BAP MGA-2024-45372); Istanbul Teknik Üniversitesi, IT; Ulusal Yüksek Başarımlı Hesaplama Merkezi, Istanbul Teknik Üniversitesi, UHeM, (4016562023, 1016682023); Ulusal Yüksek Başarımlı Hesaplama Merkezi, Istanbul Teknik Üniversitesi, UHeM
dc.description.sponsorshipInstitute of Electrical and Electronics Engineers, Signal Processing Society
dc.identifier.doi10.1109/ICIP55913.2025.11084541
dc.identifier.endpage2234
dc.identifier.isbn979-833152379-4
dc.identifier.issn1522-4880
dc.identifier.scopus2-s2.0-105028578054
dc.identifier.scopusqualityQ2
dc.identifier.startpage2229
dc.identifier.urihttps://doi.org/10.1109/ICIP55913.2025.11084541
dc.identifier.urihttps://hdl.handle.net/11411/10237
dc.indekslendigikaynakScopus
dc.language.isoen
dc.publisherIEEE Computer Society
dc.relation.ispartofProceedings - International Conference on Image Processing, ICIP
dc.relation.publicationcategoryKonferans Öğesi - Uluslararası - Kurum Öğretim Elemanı
dc.rightsinfo:eu-repo/semantics/closedAccess
dc.snmzKA_Scopus_20260402
dc.subjectChannel Splitting
dc.subjectComputational Efficiency
dc.subjectDeep Learning
dc.subjectEfficient Architecture
dc.subjectFeature Extraction
dc.subjectHigh Throughput
dc.subjectMax Pooling
dc.subjectMulti-Head Attention
dc.subjectSpatial Mixing
dc.subjectSplitter
dc.titleSplitter: Faster Inference Through Channel Partitioning and Feature Fusion
dc.typeConference Paper

Dosyalar

Koleksiyon

Scopus Indexed Publications