MSRMNet: Multi-scale skip residual and multi-mixed features network for salient object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient Object Detection (SOD) models have advanced using multi-scale feature fusion.
• Existing SOD models struggle with scale variations and produce blurred object boundaries.

Purpose of the Study:

• To develop a novel SOD model that overcomes limitations in scale detection and boundary prediction.
• To enhance the accuracy of object localization and edge detail in salient object detection.

Main Methods:

• A transformer backbone is employed to capture multi-feature layers.
• Multi-scale skip residual connections are utilized during encoding for improved positional and edge accuracy.
• Mixed feature operations in the decoding stage extract richer multi-scale semantic information.
• The Structure Similarity Index Measure (SSIM) function is incorporated into the loss function to refine boundary prediction.

Main Results:

• The proposed model achieves state-of-the-art performance on five public datasets.
• Significant improvements in performance metrics for salient object detection tasks are demonstrated.
• Enhanced accuracy in predicting object positions and precise target boundaries.

Conclusions:

• The developed transformer-based SOD model effectively addresses challenges with scale variations and boundary definition.
• The integration of multi-scale features, skip connections, and SSIM loss contributes to superior SOD performance.
• The algorithm represents a significant advancement in salient object detection technology.