Fixed source superposition and frequency partitioning limit how communication payload interacts with a probing reference in ISAC. We learn nonlinear waveform fusion across scales to enable image transmission and multi‑target localization in range and angle. Reception exploits the information available at both ends. Adaptive temporal feature weighting and spectral updates merge image features with a deterministic linear‑frequency‑modulated reference in a shared complex state, producing a waveform with fixed total energy over the array frame. At the user equipment, the reference and channel state information condition a neural demixer that reconstructs the image. At the base station, the realized transmit pulse forms a coherent dictionary for sequential localization, refined by learned local correction, physical adjustment, and residual updates. We evaluate on 512 independent test images, comparing the proposed waveforms with time‑domain superposition and learned spectral linear fusion, using a common UE receiver and a baseline sensing receiver based on coherent peak search. Averaged over eleven task weights, the proposed waveforms reduce normalized localization error by 29.0% and 27.0% respectively, while increasing mean peak signal‑to‑noise ratio by 1.40 dB and 0.82 dB. With the waveforms fixed, learned local correction further lowers mean normalized localization error by 12.2% compared to the sequential receiver without correction.
Review