MobileNetV5-Style Backbone

• MobileNetV5-style backbones are mobile-efficient vision architectures defined by hierarchical IR blocks, sparse multi-query attention, and adaptive quantization for robust low-precision inference.
• They employ a staged design integrating depthwise separable convolutions and lightweight transformer-style modules to maximize data locality on embedded accelerators.
• Empirical evaluations show significant latency reductions and quantization stability, achieving competitive accuracy with lower computational costs compared to earlier MobileNet variants.

A MobileNetV5-style backbone defines a class of mobile-efficient vision architectures characterized by hierarchical depthwise-separable inverted residual (IR) blocks, sparse attention bottlenecks, adaptive quantization mechanisms, and multi-scale fusion strategies. Unlike preceding MobileNet variants, MobileNetV5 explicitly co-designs its topology for hardware-efficient integer inference (INT4/8/16), robust quantization, and efficient multi-modal processing—including NPU-native integration for vision-language tasks. Canonical instantiations employ innovations from works such as “AutoNeural: Co-Designing Vision-LLMs for NPU Inference” (Chen et al., 2 Dec 2025), EdgeNeXt (Maaz et al., 2022), RapidNet (Munir et al., 14 Dec 2024), and LOLViT (Li et al., 2 Aug 2025). Central to the backbone is a staged architecture, integrating local inductive bias via separable convolutions and leveraging lightweight transformer-style or windowed attention modules only in late layers, thereby maximizing data locality and throughput on embedded accelerators.

1. Layer-by-Layer Stage Architecture

A typical MobileNetV5-style backbone consists of the following staged processing pipeline:

Stage	Spatial Size	Block Type	In→Out Channels
Stem	768×768	Conv2d (3×3, stride=2)	3 → C₀
Stage 1	384×384	IR Block × N₁	C₀ → C₁
Stage 2	192×192	IR Block × N₂	C₁ → C₂
Stage 3	96×96	IR × N₃ + sparse MQA	C₂ → C₃
Stage 4	48×48	IR × N₄ + sparse MQA	C₃ → C₄
MSFA	16×16	Upsample + Concat + IR + Pool	C₃+C₄ → 2048

IR blocks are formed by pointwise 1×1 convolutions followed by depthwise 3×3 convolutions and output pointwise 1×1 convolutions, with GELU activation and BatchNorm or RMSNorm depending on the stage. Sparse Multi-Query Attention (MQA) bottlenecks are integrated only in low-resolution stages (typically ≤48×48), offering lightweight context mixing without incurring quadratic memory overhead. The Multi-Scale Fusion Adapter (MSFA) aggregates outputs from the final two stages, applying upsample-concat, IR blocks, and average pooling, with flattened outputs of size 256×2048 tokens (Chen et al., 2 Dec 2025).

EdgeNeXt and RapidNet offer alternative MobileNetV5-style variations, substituting hybrid CNN-transformer stages or multi-level dilated convolutional blocks, respectively, with quantifiable gains in efficiency and accuracy on standard mobile benchmarks (Maaz et al., 2022, Munir et al., 14 Dec 2024).

2. Mathematical Formulation of Core Operations

Depthwise separable convolution is foundational, expressed as channel-wise spatial convolution followed by channel mixing: $$y_{:,:,c} = \sum_{(u,v)\in\{-1,0,1\}^2} w^{\rm dw}_{u,v,c}\ x_{\cdot + u,\, \cdot + v,\,c}$$

$$z_{i,j,k} = \sum_{c=1}^{C_{\rm in}} w^{\rm pw}_{c,k}\ y_{i,j,c}$$

Activation distributions are stabilized for integer quantization via elementwise clamping: $$a_{\text{clamped}} = \max(-\alpha,\,\min(\alpha,\,a))$$ Quantization adheres to uniform affine procedure: $$q = \operatorname{Clamp}_{[q_{\min},\,q_{\max}]} \left( \operatorname{round} \left( \frac{A}{s} \right) + z \right)$$ where the zero-point $z$ maps real zero to integer zero, and scale $s$ is derived from calibration for minimal quantization error (Chen et al., 2 Dec 2025).

In hybrid or attention-augmented MobileNetV5 instantiations, Split Depth-wise Transpose Attention (STDA) and Fast Window Attention (FWA) provide scalable context aggregation. FWA reduces full MHSA computational complexity $O(N^2 d)$ to $O(N F d)$, where $F\ll N$ via window aggregation and DReLU-based attention (Li et al., 2 Aug 2025).

3. Design Principles for Mobile-Efficient INT-x Inference

MobileNetV5 is grounded in the following design principles:

• Local-first inductive bias: Channel-wise depthwise convolutions maintain tight, bounded activation distributions and spatial locality, which are natively optimized for NPUs (Chen et al., 2 Dec 2025, Munir et al., 14 Dec 2024).
• Quantization stability: The pipeline convolution $\rightarrow$ BatchNorm/RMSNorm $\rightarrow$ GELU is empirically calibrated to clamp activations for low-precision (INT4/8/16) quantization, minimizing quantization error.
• Attention downsampling: Quadratic memory and computational costs of full self-attention are avoided either by eliminating global MHSA in early stages, aggregating only in low-resolution stages (sparse MQA, FWA), or replacing with lightweight local-global fusions (Maaz et al., 2022, Li et al., 2 Aug 2025).
• Elimination of dynamic statistics: Any architectural feature introducing sample-dependent statistics (e.g., squeeze-and-excitation, dynamic gating) is excised to allow consistent static range calibration for integer quantization (Chen et al., 2 Dec 2025).
• Multi-scale feature fusion: Multi-scale fusion modules compress hierarchical information, employing upsample/concat and IR blocks, or multi-level dilated convolutions for expanded receptive field (Munir et al., 14 Dec 2024).

4. Quantitative Efficiency and Performance Metrics

Empirical evaluation of MobileNetV5-style backbones demonstrates significant efficiency improvements over both transformer-based and hybrid models:

Model	Params	MACs	Latency(ms)	Top-1 (%)
MobileNetV2×1.4	6.1 M	0.6 G	1.1	74.7
RapidNet-Ti	6.6 M	0.6 G	0.9	76.3
EdgeNeXt-S	5.6 M	1.3 G	–	79.4
MobileNetV5	300 M	35 G	278.1 (768²)	–

On a Qualcomm SA8295P NPU, MobileNetV5-style encoders delivered a 5.8×–14× latency reduction over InternViT, with quantization error reduced by 7× (SQNR of 45 dB vs 28 dB for W8A16 quantization) (Chen et al., 2 Dec 2025). RapidNet-Ti achieves superior speed and accuracy compared to MobileNetV2×1.4 and hybrid baselines (Munir et al., 14 Dec 2024). EdgeNeXt surpasses MobileNetV3 and MobileNetV2 at equivalent parameter counts on ImageNet (Maaz et al., 2022). LOLViT with FWA realizes a 5× speedup over MobileViT-X while maintaining comparable accuracy on COCO and BDD100K segmentation (Li et al., 2 Aug 2025).

5. Integration of Sparse Attention, Multi-Scale Fusion, and Dilated Convolutions

Sparse MQA bottlenecks are allocated to late stages (e.g., 96×96/48×48) to capture cross-token dependencies with minimal parameter overhead. In EdgeNeXt, the STDA block splits input channels, mixes features via depthwise convolution, and computes cross-channel attention, culminating in efficient multi-scale encoding (Maaz et al., 2022). RapidNet's Multi-Level Dilated Convolution (MLDC) combines parallel dilated depthwise convolutions ($d=2$, $d=3$) and large-kernel FFN paths to spatially aggregate context over 5×5, 7×7, and reparameterized 7×7 domains, yielding expanded theoretical receptive fields (Munir et al., 14 Dec 2024). LOLViT's Fast Window Attention leverages adaptive key/value reduction via window aggregation and DReLU, tightly controlling sequence length and computational saturation in global-local fusion bottlenecks (Li et al., 2 Aug 2025).

6. Comparison with Preceding MobileNet and Hybrid Architectures

MobileNetV5 diverges from MobileNetV2/V3 by:

• Replacing squeeze-and-excitation blocks and dynamic gating with static, clamped distributions for reliable quantization.
• Reducing quadratic attention computation via windowed or sparse aggregation schemes, only in late network stages.
• Incorporating multi-scale fusion adapters and multi-level dilated convolutions to strengthen spatial feature mixing for detection/segmentation tasks.
• Achieving higher ImageNet accuracy at lower or comparable parameter counts (EdgeNeXt-S: 79.4% @5.6M params vs MobileNetV2: 74.7% @6.9M params) (Maaz et al., 2022).
• Demonstrating state-of-the-art detection (RapidNet-M: APbox=42.0, APmask=38.3 @17.3M vs Swin-Tiny: APbox=42.2, APmask=39.1 @29.0M) (Munir et al., 14 Dec 2024)
• Realizing robust integer-only inference, multi-modal fusion, and longer effective context windows for vision-LLMs on edge hardware (Chen et al., 2 Dec 2025).

7. Guidelines, Ablations, and Deployment Considerations

Designing a MobileNetV5 backbone requires:

• Adaptive window size in attention: Window size $P_w$ is selected such that attention key/value sequence is reduced by $P_w^2$, optimizing latency-accuracy tradeoffs (Li et al., 2 Aug 2025).
• Stage-wise allocation: Insert sparse attention/fusion blocks only where spatial size $H\times W \leq 16\times16$.
• Embedding/channel expansion: Expansion factors (e.g., $c_{\rm exp}=4$ in IR blocks) facilitate stable feature fusion.
• Quantization calibration: Static clamping thresholds $\alpha$ determined via layer-wise calibration guarantee error bounds on INT4/8/16 deployment (Chen et al., 2 Dec 2025).
• Key-caching and window folding: Key/value aggregation can be further downsampled to minimize recomputation, with optional folding for deeper stacking (Li et al., 2 Aug 2025).
• Ablation findings: Multi-level dilation, large-kernel FFNs, and local-global convolutional fusion each contribute measurable gains in accuracy with marginal compute increase; dynamic attention variants (FWA, STDA, sparse MQA) avoid saturation and memory constraints characteristic of classical MHSA (Maaz et al., 2022, Munir et al., 14 Dec 2024, Li et al., 2 Aug 2025).

MobileNetV5-style backbones thus define a modular, hardware-oriented design paradigm, delivering scalable accuracy and throughput for mobile and embedded vision deployments, and setting the foundation for next-generation NPU-native multi-modal models.