Head-mounted augmented-reality (AR) devices are
increasingly used by first responders and military medics to
visualize radio-frequency (RF) tracks, casualty vitals and threat
signatures in real time. These platforms operate under severe
resource constraints: the computational budget is on the order
of tens of milliseconds, the power budget is under one watt, and
the thermal headroom is limited by the user’s skin. Prior work
demonstrated that RF–AR situational awareness can be achieved
within ∼200 ms end-to-end on uncompressed networks. However,
the neural networks used for classification and localization are
heavily over-parameterized, leading to energy-intensive inference
and lengthy stalls on battery-powered wearables. To tackle this
problem, we present a pipeline for on-device RF filtering and
compression that combines quantization, sparsity and knowledge
distillation to shrink models without compromising mission utility. Quantization reduces the precision of weights and activations,
lowering memory footprints and enabling faster integer arithmetic [1], while magnitude-based pruning removes unimportant
parameters and accelerates inference [2]. Recent studies show
that pruning and quantization jointly diminish computational
and memory requirements [3] but must be applied carefully
because their effects are non-orthogonal [4]. We further employ
teacher–student knowledge distillation, transferring knowledge
from a high-capacity ”teacher” network to a lightweight ”student” model [5], [6]. Our experiments on Jetson-class edge devices
and Pixel-8 smartphones sweep multiple quantization bit-widths
and sparsity levels, producing accuracy–latency–power Pareto
curves. At 50 ms median latency and 0.9 W average power, our
distilled INT8/70 % sparse student attains within 1 % of baseline
accuracy, yielding >5× energy savings. Hardware-aware model
compression techniques [7] and adaptive bit-width selection [8]
enable deployment on resource-constrained wearable platforms.
We release our code, datasets and measurement harness to foster
reproducible research in RF–AR compression.Head-mounted augmented-reality (AR) devices are
increasingly used by first responders and military medics to
visualize radio-frequency (RF) tracks, casualty vitals and threat
signatures in real time. These platforms operate under severe
resource constraints: the computational budget is on the order
of tens of milliseconds, the power budget is under one watt, and
the thermal headroom is limited by the user’s skin. Prior work
demonstrated that RF–AR situational awareness can be achieved
within ∼200 ms end-to-end on uncompressed networks. However,
the neural networks used for classification and localization are
heavily over-parameterized, leading to energy-intensive inference
and lengthy stalls on battery-powered wearables. To tackle this
problem, we present a pipeline for on-device RF filtering and
compression that combines quantization, sparsity and knowledge
distillation to shrink models without compromising mission utility. Quantization reduces the precision of weights and activations,
lowering memory footprints and enabling faster integer arithmetic [1], while magnitude-based pruning removes unimportant
parameters and accelerates inference [2]. Recent studies show
that pruning and quantization jointly diminish computational
and memory requirements [3] but must be applied carefully
because their effects are non-orthogonal [4]. We further employ
teacher–student knowledge distillation, transferring knowledge
from a high-capacity ”teacher” network to a lightweight ”student” model [5], [6]. Our experiments on Jetson-class edge devices
and Pixel-8 smartphones sweep multiple quantization bit-widths
and sparsity levels, producing accuracy–latency–power Pareto
curves. At 50 ms median latency and 0.9 W average power, our
distilled INT8/70 % sparse student attains within 1 % of baseline
accuracy, yielding >5× energy savings. Hardware-aware model
compression techniques [7] and adaptive bit-width selection [8]
enable deployment on resource-constrained wearable platforms.
We release our code, datasets and measurement harness to foster
reproducible research in RF–AR compression.