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

We present a decoupled multi-task framework trained via a two-stage scheme and augmented with channel attention for optical image compression-reconstruction and classification. In Stage 1 (classification-focused pretraining), the encoder is optimized without quantization noise to learn stable, semantically discriminative representations. In Stage 2, quantization and decoding are enabled, and channel attention is inserted to emphasize high-information-carrying channels, preserving key structures and textures under compression. The design exploits shallow low-level cues (edges and textures) for reconstruction and deep high-level semantics (object structure and category) for classification, yielding complementary cross-task benefits. Experiments demonstrate a robust trade-off: across bit-per-pixel (BPP) settings, classification accuracy remains at 81.40%, surpassing a joint MTL baseline (60.20% at <i>B</i><i>P</i><i>P</i>=0.0362) and a JPEG2000 + VGG19 pipeline (20.80%). The model uses only 1.97×10⁷ parameters-over <b>33%</b> fewer than alternatives-and achieves nearly ×3 faster training under identical settings. In sum, the proposed approach jointly ensures classification accuracy, reconstruction quality, and computational efficiency, delivering high performance, a lightweight footprint, and strong deployability for resource-constrained, low-bandwidth edge scenarios.