Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

DFIR-DETR: Frequency-domain iterative refinement and dynamic feature aggregation for small object detection.

Bo Gao1, Jingcheng Tong1, Xingsheng Chen2

  • 1School of Information Engineering, Beijing Institute of Graphic Communication, Beijing, 102600, China.

Neural Networks : the Official Journal of the International Neural Network Society
|June 17, 2026
PubMed
Summary

Related Concept Videos

Linear Approximation in Frequency Domain01:26

Linear Approximation in Frequency Domain

Linear systems are characterized by two main properties: superposition and homogeneity. Superposition allows the response to multiple inputs to be the sum of the responses to each individual input. Homogeneity ensures that scaling an input by a scalar results in the response being scaled by the same scalar.
In contrast, nonlinear systems do not inherently possess these properties. However, for small deviations around an operating point, a nonlinear system can often be approximated as linear.

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Dimension-controlled formation of crease patterns on soft solids.

Soft matter·2016
Same author

Modeling Day-to-day Flow Dynamics on Degradable Transport Network.

PloS one·2016
Same author

Tetramethylpyrazine Protects Against Glucocorticoid-Induced Apoptosis by Promoting Autophagy in Mesenchymal Stem Cells and Improves Bone Mass in Glucocorticoid-Induced Osteoporosis Rats.

Stem cells and development·2016
Same author

Corrigendum: Lithium-ion-based solid electrolyte tuning of the carrier density in graphene.

Scientific reports·2016
Same author

PTEN/PI3K/AKT protein expression is related to clinicopathological features and prognosis in breast cancer with axillary lymph node metastases.

Human pathology·2016
Same author

Comparing the Diagnostic Accuracy of RTE and SWE in Differentiating Malignant Thyroid Nodules from Benign Ones: a Meta-Analysis.

Cellular physiology and biochemistry : international journal of experimental cellular physiology, biochemistry, and pharmacology·2016
This summary is machine-generated.

DFIR-DETR enhances small object detection by addressing uniform attention, norm drift, and spatial filtering issues in neural networks. This novel approach improves accuracy on complex datasets with fewer parameters and computations.

Area of Science:

  • Computer Vision
  • Deep Learning
  • Object Detection

Background:

  • Neural network design faces challenges in small object detection within complex scenes.
  • Existing models like RT-DETR exhibit uniform attention, norm drift during upsampling, and spatial filtering that degrades high-frequency details crucial for small objects.

Purpose of the Study:

  • To develop an improved object detection model, DFIR-DETR, specifically targeting the limitations of the RT-DETR baseline for small object detection.
  • To address deficiencies in attention mechanisms, feature upsampling normalization, and convolutional filtering that hinder performance on small objects.

Main Methods:

  • DFIR-DETR was developed by analyzing and modifying specific modules of the RT-DETR architecture.
  • Key modifications include addressing uniform attention, compensating for norm drift in pyramid necks, and refining bottleneck convolutions to preserve high-frequency components.
Keywords:
Cross-scene generalisationFrequency-domain feature learningMulti-scale feature fusionSmall object detectionSparse attention mechanismTransformer-based detection

Related Experiment Videos

Main Results:

  • DFIR-DETR achieved 92.9% mAP50 on the NEU-DET dataset and 51.6% mAP50 on the VisDrone dataset.
  • The model demonstrates significant performance gains with only 11.7M parameters and 47.2 GFLOPs, indicating high efficiency.

Conclusions:

  • DFIR-DETR effectively overcomes the limitations of previous models in detecting small objects in complex environments.
  • The proposed method offers a computationally efficient and accurate solution for small object detection across diverse visual domains.