VeloxNet: Efficient Spatial Gating for Lightweight Embedded Image Classification

Md Meftahul Ferdaus; Elias Ioup; Mahdi Abdelguerfi; Anton Netchaev; Steven Sloan; Ken Pathak; Kendall N. Niles

cs.CVMar 19, 2026

VeloxNet: Efficient Spatial Gating for Lightweight Embedded Image Classification

Authors: Md Meftahul Ferdaus, Elias Ioup, Mahdi Abdelguerfi, Anton Netchaev, Steven Sloan, Ken Pathak, Kendall N. Niles

Organizations: Canizaro Livingston Gulf States Center for Environmental Informatics, the University of New Orleans, New Orleans, USA · Center for Geospatial Sciences, Naval Research Laboratory, Stennis Space Center, Hancock County, Mississippi, USA · US Army Corps of Engineers, Engineer Research and Development Center, Vicksburg, Mississippi, USA

Abstract

Deploying deep learning models on embedded devices for tasks such as aerial disaster monitoring and infrastructure inspection requires architectures that balance accuracy with strict constraints on model size, memory, and latency. This paper introduces VeloxNet, a lightweight CNN architecture that replaces SqueezeNet's fire modules with gated multi-layer perceptron (gMLP) blocks for embedded image classification. Each gMLP block uses a spatial gating unit (SGU) that applies learned spatial projections and multiplicative gating, enabling the network to capture spatial dependencies across the full feature map in a single layer. Unlike fire modules, which are limited to local receptive fields defined by small convolutional kernels, the SGU provides global spatial modeling at each layer with fewer parameters. We evaluate VeloxNet on three aerial image datasets: the Aerial Image Database for Emergency Response (AIDER), the Comprehensive Disaster Dataset (CDD), and the Levee Defect Dataset (LDD), comparing against eleven baselines including MobileNet variants, ShuffleNet, EfficientNet, and recent vision transformers. VeloxNet reduces the parameter count by 46.1% relative to SqueezeNet (from 740,970 to 399,366) while improving weighted F1 scores by 6.32% on AIDER, 30.83% on CDD, and 2.51% on LDD. These results demonstrate that substituting local convolutional modules with spatial gating blocks can improve both classification accuracy and parameter efficiency for resource-constrained deployment. The source code will be made publicly available upon acceptance of the paper.

Figures & tables

Fig. 1: Macro-architectural view of the original SqueezeNet architecture along with micro-architectural view showing organization of convolution filters in the fire module

Fig. 2: Macro-architectural view of the proposed VeloxNet architecture along with micro-architectural view of the gMLP architecture featuring our efficient SGU implementation

SqueezeNet			VeloxNet
Layer name/type	Output size	Number of parameters	Layer name/type	Output size	Number of parameters
input image	$224\times 224\times 3$	-	input image	$224\times 224\times 3$	-
conv1 (7 $\times$ 7)	$112\times 112\times 96$	14,112	conv1 (3 $\times$ 3)	$111\times 111\times 156$	4,212
batchnorm1	$112\times 112\times 96$	192	batchnorm1	$111\times 111\times 156$	6
maxpool1	$56\times 56\times 96$	-	maxpool1	$55\times 55\times 156$	-
fire2	$56\times 56\times 128$	12,064	gMLP2	$55\times 55\times 156$	49,296

TABLE I: Architectural dimensions: SqueezeNet vs. VeloxNet

Dataset	Class	Train	Validation	Testing	Total per Class
AIDER	Collapsed Building	367	41	103	511
	Fire	249	63	209	521
	Flood	252	63	211	526
	Traffic	232	59	194	485
	Normal	2107	527	1756	4390
	Total Per Set	3207	753	2473	6433

TABLE II: Distribution of images for each class within the training, validation, and testing sets across the AIDER, CDD, and LDD datasets

Methods	F1	Precision	Recall	Parameters	FLOPs	Size	FPS
	(%)	(%)	(%)	( $\downarrow$ )	(M)	(MB)	( $\uparrow$ )
MobileNetV1	73.13	74.13	73.20	3,212,101	588	12.34	411.30
MobileNetV2	72.94	73.56	73.01	2,230,277	326	8.64	213.48
MobileNetV3	63.96	65.46	63.88	3,660,573	233	14.06	160.52
ShuffleNet	73.45	73.73	73.59	1,258,729	152	4.86	175.96
EfficientNet	67.60	69.08	68.16	3,812,005	396	14.69	152.39

TABLE III: Performance comparison of VeloxNet and state-of-the-art lightweight CNNs on AIDER dataset

Methods	F1	Precision	Recall	Parameters	FLOPs	Size	FPS
	(%)	(%)	(%)	( $\downarrow$ )	(M)	(MB)	( $\uparrow$ )
MobileNetV1	62.26	64.28	62.43	3,212,101	590	12.34	426.02
MobileNetV2	61.33	61.29	61.52	2,230,277	330	8.64	230.79
MobileNetV3	52.68	53.11	53.08	3,660,573	230	14.06	169.88
ShuffleNet	60.65	61.25	60.99	1,258,729	150	4.86	177.18
EfficientNet	65.54	66.49	66.31	3,812,005	400	14.69	151.58

TABLE IV: Performance comparison of VeloxNet and state-of-the-art lightweight CNNs on CDD dataset

Methods	F1	Precision	Recall	Parameters	FLOPs	Size	FPS
	(%)	(%)	(%)	( $\downarrow$ )	(M)	(MB)	( $\uparrow$ )
MobileNetV1	82.71	83.58	82.94	3,212,101	590	12.34	414.29
MobileNetV2	82.77	83.42	82.94	2,230,277	330	8.64	226.12
MobileNetV3	77.04	77.44	77.25	3,660,573	230	14.06	163.09
ShuffleNet	80.95	81.66	81.18	1,258,729	150	4.86	182.29
EfficientNet	77.67	81.01	78.04	3,812,005	400	14.69	155.68

TABLE V: Performance comparison of VeloxNet and state-of-the-art lightweight CNNs on LDD dataset

Variant	F1 (%)	Params	Modification
Full model
VeloxNet (full)	81.57	399K	Proposed architecture
SqueezeNet (fire modules)	75.25	741K	All gMLP blocks replaced with fire modules
Component removal
w/o SGU	77.83	385K	Remove spatial gating; keep channel projections only
w/o residual connections	74.19	399K	Remove skip connections from gMLP blocks

TABLE VI: Ablation study on the AIDER dataset. Each row modifies one aspect of the full VeloxNet architecture.

Image Classification

Image classification assigns semantic labels to images, enabling recognition of objects, scenes, conditions, and fine-grained categories.

See all 391 papers

When Less Is More: A Controlled Benchmark of Lightweight CNNs for Satellite Land-Cover Segmentation on DeepGlobe
Jul 25, 2026Atiq Ur Rehman, Joseph Michael DonovanRemote Sensing Image Segmentation Land Cover Classification
Department of Economics & Decision Sciences, University of South Dakota, Vermillion, SD, USA · University of South Dakota, Vermillion, USA
EdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI
Jul 8, 2026Hao Kong, Di Liu, Shuo Huai +5Edge Inference Convolutional Neural Networks
School of Computer Science and Engineering, Nanyang Technological University (NTU), Singapore · HP-NTU Digital Manufacturing Corporate Lab, Nanyang Technological University (NTU), Singapore · Department of Computer Science, Norwegian University of Science and Technology (NTNU), Trondheim, Norway +1
UtVAA: Ultra-tiny Vision Transformer with Affix Attention for Mobile Image Classification
Jun 2, 2026Romiyal George, Sathiyamohan Nishankar, Selvarajah Thuseethan +1Vision Transformer Edge Inference
University of Peradeniya, Sri Lanka · Charles Darwin University, Australia

Jul 25, 2026J/K move · Enter open · S save

When Less Is More: A Controlled Benchmark of Lightweight CNNs for Satellite Land-Cover Segmentation on DeepGlobe

Atiq Ur Rehman, Joseph Michael Donovan

Department of Economics & Decision Sciences, University of South Dakota, Vermillion, SD, USA · University of South Dakota, Vermillion, USA

Remote Sensing Image Segmentation Land Cover Classification Remote Sensing Convolutional Neural Networks Remote Sensing Benchmarks Lightweight CNNs

Open paper AskPDF arXiv

High-resolution satellite imagery is the backbone of good land-cover classification, and without that, environmental monitoring, urban planning, and sustainable resource management all fall short. Deep learning architectures perform well in semantic segmentation, but the efficiency-accuracy trade-off across classical convolutional encoders is not well quantified under controlled, reproducible conditions. This study compares five architectures VGG16, MobileNetV2, InceptionV3, AlexNet, and CNN on the DeepGlobe Land Cover Classification dataset using three progressively optimized iterations to isolate regularisation, transfer learning, and architectural depth. To ensure performance differentials reflect architectural properties, all experiments used identical preprocessing, hyperparameter, and training protocols without data augmentation or class-imbalance correction. At 24.98 MB, MobileNetV2_v1 had the highest overall accuracy (0.7906) and mean Intersection over Union (0.4625), outperforming deeper alternatives like InceptionV3_v2 (125.17 MB, accuracy 0.7610) and VGG16_v2 (71.13 MB, accuracy 0.7653). Class-wise analysis showed strength in urban, agricultural, and water categories, but rangeland-barren confusion showed that architectural optimization alone cannot optimize spectrally similar minority classes. Strong spatial generalization and crisp boundary delineation were confirmed on held-out test imagery, validating operational applicability. These results show that lightweight, transfer-learned models can match or outperform deeper models in resource-constrained remote-sensing environments, enabling scalable land-cover mapping.

Abstract

Image Classification

Explore similar work

When Less Is More: A Controlled Benchmark of Lightweight CNNs for Satellite Land-Cover Segmentation on DeepGlobe

EdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI

UtVAA: Ultra-tiny Vision Transformer with Affix Attention for Mobile Image Classification