Saliency Segmentation Oriented Deep Image Compression with Novel Bit Allocation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image compression can degrade machine analysis task performance.
• Deep learning methods for image compression are advancing, but lack focus on saliency segmentation.
• Existing methods often couple compression and analysis networks, limiting compatibility.

Purpose of the Study:

• To develop a deep image compression method specifically optimized for saliency segmentation.
• To decouple compression and segmentation networks for broader applicability.
• To improve the rate-accuracy trade-off in salient object detection.

Main Methods:

• Proposing a deep compression network that prioritizes local signal fidelity for salient pixels.
• Implementing a bit allocation strategy using probability distributions and an Ascending Cosine Roll-Down (ACRD) function.
• Training the compression network independently, decomposing latent representations into base and enhancement channels.

Main Results:

• Achieved an average bitrate saving of 10.34% compared to state-of-the-art methods.
• Demonstrated improved rate-accuracy (R-A) performance on sixteen downstream saliency segmentation networks.
• Validated effectiveness across five conventional Salient Object Detection (SOD) datasets.

Conclusions:

• The proposed method effectively enhances saliency segmentation by intelligently allocating bits to important image regions.
• Decoupling compression and segmentation networks improves compatibility with diverse saliency models.
• This approach offers a significant advancement in optimizing image compression for machine perception tasks like saliency segmentation.