Abstract
Unmanned aerial vehicle (UAV) control signals are often bursty or utilize frequency hopping, so the evidence available in a spectrogram varies over time. We study confidence-adaptive multi-segment early exit (CAMSE) and separately evaluate online statistical adaptation of batch normalization (OSA-BN). CAMSE sequentially processes temporal segments with a shared ResNet18 and stops when the top-two probability margin from averaged logits reaches a stage-specific threshold. Temporal-segment augmentation supports segment-wise inference. An eight-class RFUAV evaluation separates 82 source IQ files into training, tuning, and test partitions. At signal-to-noise ratios (SNRs) from −30 to −22 dB, CAMSE achieves 48.53% accuracy with 2.47 segments per sample, compared with 49.07% for fixed-five fusion, reducing backbone evaluations by 50.51%. Regular-SNR accuracy is 99.52% with 1.06 segments. The whole-image Frozen-BN reference achieves 60.53% extreme-SNR accuracy. Uniform and decreasing thresholds offer different accuracy–computation operating points. In the separate whole-image experiment, OSA-BN raises extreme-SNR accuracy to 76.13% without labels, SNR estimates, or gradient updates. On eight native-noise IQ windows, CAMSE reduces single-thread CPU latency from 498.41 to 378.97 ms relative to fixed-five inference. These results support adaptive stopping within multi-segment processing and statistical adaptation for whole-image recognition.