Hybrid aggregation strategy with double inverted residual blocks for lightweight salient object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Lightweight salient object detection (SOD) is crucial for efficient real-time applications.
• Hybrid encoders combining CNNs and transformers offer improved accuracy-efficiency balance for SOD.
• Existing methods struggle with effective feature aggregation from hybrid encoders in lightweight SOD.

Purpose of the Study:

• To propose a novel Hybrid Aggregation Strategy Network (HASNet) for efficient and accurate lightweight salient object detection.
• To address the challenge of aggregating features from CNNs and transformers in hybrid SOD models.
• To balance accuracy and efficiency in lightweight SOD models.

Main Methods:

• Developed HASNet utilizing a hybrid aggregation strategy for CNN and transformer features.
• Implemented deep aggregation with Global Inverted Residual Blocks (GIRB) for transformer features.
• Used Lightweight Inverted Residual Blocks (LIRB) for shallow aggregation of convolutional features.

Main Results:

• HASNet demonstrated superior performance across five datasets in terms of accuracy, speed, and parameter size.
• The proposed GIRB and LIRB effectively facilitated cross-architecture feature transfer and fusion.
• The model achieved a significant improvement over existing lightweight SOD methods.

Conclusions:

• HASNet offers an effective solution for lightweight salient object detection by intelligently aggregating hybrid features.
• The network achieves a strong balance between computational efficiency and detection accuracy.
• The findings suggest a promising direction for future research in efficient deep learning models for computer vision tasks.