Transformer with difference convolutional network for lightweight universal boundary detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning excels at boundary detection but requires large models and datasets, leading to high computational costs.
• Existing methods often lack generalizability across different datasets, necessitating dataset-specific retraining.
• A need exists for efficient, universal boundary detection models with fewer parameters.

Purpose of the Study:

• To develop a lightweight, universal boundary detection method using a hybrid convolution-transformer architecture.
• To investigate a single model's ability to perform cross-dataset boundary detection without retraining.
• To reduce computational power consumption in deep learning-based boundary detection.

Main Methods:

• Introduced the Transformer with Difference Convolutional Network (TDCN), integrating convolution and transformer components.
• Employed a convolution network with edge operators for multiscale difference feature extraction.
• Developed a boundary-aware self-attention mechanism within the transformer and an attention loss function incorporating boundary direction.

Main Results:

• The TDCN achieved competitive performance on multiple public datasets with significantly fewer model parameters.
• Demonstrated universal prediction capabilities, performing effectively on unseen datasets without retraining.
• Validated the model's effectiveness in achieving high-performance, low-parameter boundary detection.

Conclusions:

• The proposed TDCN offers an efficient and effective solution for universal boundary detection.
• The hybrid architecture and boundary-aware attention mechanism contribute to improved cross-dataset generalization.
• This research advances low-level vision tasks by enabling computationally inexpensive, adaptable boundary detection models.