Abstract
No public respiratory sound dataset provides event-level time-frequency mask annotation, so the localization of crackles and wheezes must be learned under weak supervision. This study evaluates whether two architectural priors from camouflaged object detection, frequency attention and reverse attention, transfer to this task. They are implemented as optional modules on a U-Net and tested in two studies: a synthetic benchmark with exact masks and controlled camouflage, and the SPRSound database with full-band annotations over ten patient-disjoint splits. The benchmark shows that the broadband signal-to-noise ratio does not measure camouflage: at the same nominal value, a wheeze sits 26 dB above a crackle within its own mask. Frequency attention improves event-based F1 by 0.053 over a capacity-matched control, with a positive sign in 9 of 10 splits. Reverse attention does not replicate on new seeds. The advantage attenuates on the independent test set, the drop lies in precision, and a conventional convolutional recurrent neural network is numerically superior. Five hypotheses about the mechanism are evaluated and none is confirmed. This work does not propose a competitive method: it measures what transfers and what does not.