Explainability Enhanced Object Detection Transformer With Feature Disentanglement.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Explainability is crucial for deploying deep learning models in critical applications.
• Existing object detection models, particularly DETR variants, suffer from entangled features, hindering interpretability.
• The regression function in object detection contributes to feature entanglement and reduced semantic coverage.

Purpose of the Study:

• To enhance the explainability of end-to-end object detection with Transformer (DETR) models.
• To introduce a feature disentanglement method to improve model interpretability and performance.
• To address the limitations of entangled features in deep learning-based object detection.

Main Methods:

• A divide-and-conquer decoupling paradigm was employed for feature learning.
• Tensor Singular Value Decomposition (T-SVD) was utilized to generate feature bases.
• Batch averaged Feature Spectral Penalization (BFSP) loss was introduced to constrain feature disentanglement and balance semantic activation.

Main Results:

• The proposed Explainability Enhanced object detection Transformer with feature Disentanglement (DETD) model demonstrated improved object detection performance.
• Consistent outperformance was observed across various backbones and DETR variants on two datasets.
• Grad-CAM visualizations confirmed enhanced feature learning explainability through feature disentanglement.

Conclusions:

• The DETD model effectively disentangles features, leading to enhanced explainability in object detection.
• The proposed method improves both detection performance and feature interpretability.
• This work offers a pathway for more trustworthy and interpretable deep learning models in critical domains.