Decoupled two-stage multi-task learning with channel attention for optical image compression-reconstruction and...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Efficient optical image compression and reconstruction are crucial for data transmission and storage.
• Simultaneously performing image classification alongside compression presents a significant challenge due to differing data requirements.

Purpose of the Study:

• To develop a decoupled multi-task learning framework for joint optical image compression-reconstruction and classification.
• To enhance representation learning for both tasks by leveraging shallow low-level cues and deep high-level semantics.

Main Methods:

• A two-stage training scheme with channel attention was employed for the multi-task framework.
• Stage 1 focused on classification-pretrained encoder optimization without quantization noise.
• Stage 2 enabled quantization and decoding, incorporating channel attention to preserve critical image structures.

Main Results:

• The framework achieved a classification accuracy of 81.40% across various bit-per-pixel (BPP) settings.
• This surpasses joint multi-task learning (MTL) baselines (60.20%) and traditional pipelines (20.80%).
• The model utilizes 33% fewer parameters and trains nearly 3 times faster than alternatives.

Conclusions:

• The proposed approach effectively balances classification accuracy, reconstruction quality, and computational efficiency.
• It offers a lightweight and deployable solution for resource-constrained, low-bandwidth edge scenarios.
• The framework demonstrates complementary cross-task benefits by exploiting both low-level and high-level image features.