Quaternion Approximate Networks for Enhanced Image Classification and Oriented Object Detection

AI summary

Overview

• Research area: Computer vision with rotation-aware deep learning. The paper targets image classification and oriented (rotated) object detection, motivated by robotic manipulation and assembly.
• Technical level: Advanced. The work involves quaternion algebra, Hamilton product decomposition, custom CUDA kernels, and modifications to YOLOv11 architecture blocks.
• Scope: The paper introduces QUAN (Quaternion Approximate Networks), a framework that approximates quaternion convolution with real-valued operations to gain rotation awareness at reduced parameter cost, and benchmarks it on classification, oriented detection, and a robotic workspace dataset.

What This Paper Is About

Standard convolutional networks and mainstream detectors rely on axis-aligned bounding boxes, which discard the object orientation that robotic grasping and assembly depend on. Quaternion neural networks can represent 3D rotations compactly and without gimbal lock, but pure quaternion implementations are poorly supported in deep learning frameworks and computationally expensive. QUAN's goal is to keep the rotational benefits of quaternion algebra while running efficiently on ordinary tensor operations and custom CUDA kernels, then carry that into both image classification and oriented object detection.

Key Contributions

1. A Hamilton-based approximation of quaternion convolution. Rather than computing all 16 terms of the Hamilton product, QUAN decomposes the operation into separable real-valued convolutions, reducing computational complexity from O(16CHW) to O(4CHW) while preserving the quaternion sign structure. Custom CUDA kernels fuse the operations, and the backward pass uses the same separable structure via a mixing matrix.
2. Two quaternion-aware architectural components. Independent Quaternion Batch Normalization (IQBN) normalizes each quaternion component separately while retaining cross-component relationships, and Quaternion Partial Spatial Attention (QC2PSA) extends spatial attention into the quaternion domain.
3. A systematic quaternion conversion of YOLOv11 blocks. The paper adapts C3k2 (as QC3k2), SPPF (as QSPPF), and C2PSA (as QC2PSA) into quaternion counterparts and replaces standard convolutions with approximated quaternion convolution.
4. An orientation-aware loss function. The total loss combines classification loss, CIoU box regression, a quaternion angular loss based on the geodesic angle between predicted and target orientations, a unit-quaternion regularization term, and an orientation smoothness term between sequential predictions.

Main Findings

• Parameter-efficient classification: QUAN-WRN reaches 95.12% on CIFAR-10 with 717K parameters, versus 94.76% for WRN-16-4 at 2.75M parameters. The paper describes this as 74% fewer parameters than WRN-16-4 while outperforming all quaternion baselines.
• CIFAR-100 and ImageNet: QUAN reports 76.83% on CIFAR-100 and 74.28% on the ImageNet subset, compared with 76.09% and 71.09% for WRN-16-4. The ImageNet figures carry a footnote: 21.8M parameters for ResNet34 and 5.39M for QUAN on ImageNet, which the paper describes as a 75% reduction.
• Quaternion baselines: Deep QCNN scores 94.56% on CIFAR-10 and 73.99% on CIFAR-100 with 932.8K parameters; QVGG11 scores 85.15% on CIFAR-10 with 3.8M parameters; Shallow QCNN scores 77.78% on CIFAR-100; QResNet34 is reported at 92–94% on CIFAR-10 with 1M parameters.
• Modest training overhead: Training time increased by 6.8% on average across classification experiments, attributed to the approximation approach. The authors also report that implementing pure quaternion operations (without their approximation) was 3.8× slower due to lack of hardware optimization for quaternion arithmetic.
• Axis-aligned detection on the robotic dataset: QUAN-YOLO11n uses 0.67M parameters versus 2.61M for YOLO11n, with mAP50 of 0.989 versus 0.992 and mAP50-95 of 0.955 versus 0.951. The text states QUAN uses only 25.3% of the parameters of YOLO11n; the table lists inference time as 1.95 ms for QUAN and 2.098 ms for YOLO11n, while the text describes QUAN as having 36% more inference time.
• Axis-aligned detection on COCO: QUAN underperforms YOLO11n, with mAP50-95 of 0.339 versus 0.3908 and mAP50 of 0.478 versus 0.547, at 1.187M versus 2.65M parameters. The paper attributes this to impaired generalization as dataset scale grows and to smaller model size, and notes the authors lacked computational capacity to match the original training batch size and hyperparameters.
• Oriented detection on the robotic dataset: QUAN-YOLO11n achieves 76.4 mAP@50-95 versus 68.5 for YOLO11n-obb, at 0.69M versus 2.66M parameters, with inference reported as 3.44 ms versus 3.57 ms (a 3.7% reduction). mAP@50 was 92.4 for QUAN versus 93.3 for YOLO11n-obb.
• Oriented detection on DOTA-v1.0: QUAN reaches 76.2 mAP@50 and 60.8 mAP@50-95 versus 77.3 and 61.4 for YOLO11n-obb, at 1.24M versus 2.66M parameters. The paper again attributes the gap to small model size.
• Ablation on normalization and mapping: Poincaré mapping generally outperformed the alternatives across activation functions, and SiLU consistently beat ReLU and QPReLU variants. Traditional quaternion batch normalization showed on average 0.61% less accuracy than IQBN across CIFAR-10 experiments.
• Strict quaternion constraints hurt: Enforcing the unit-quaternion constraint through Riemannian optimization degraded performance, which the authors cite as justification for relaxing it during training.

Methodology in Plain English

Images are treated as quaternions, four-component numbers that extend complex numbers and are naturally suited to describing 3D rotations. The authors convert RGB pixels into quaternion values using a Poincaré mapping, which projects colors onto the unit quaternion sphere so that relative color relationships are preserved.

Rather than multiplying quaternion-valued kernels and features in full, which involves 16 interaction terms, QUAN splits the operation into four ordinary convolutions combined with the correct plus and minus signs. This keeps the algebraic sign pattern that gives quaternion convolution its rotational character while running on standard, well-optimized tensor operations. Custom CUDA kernels fuse the pieces together to limit memory traffic, and the gradients for inputs and weights are computed with a 4×4 mixing matrix that mirrors the forward pass.

The network normalizes each quaternion component independently with its own learned scale and shift, which stabilizes training without forcing quaternions to be treated as single indivisible units. The authors then rebuild YOLOv11's standard blocks, the feature-mixing C3k2 module, the spatial pyramid pooling block, and the spatial attention block, in quaternion form and swap in quaternion convolutions throughout.

For oriented detection, the loss adds a term that measures the angular distance between predicted and ground-truth quaternions, a term that pushes predictions toward unit length, and a smoothing term that discourages jitter between successive predictions, all on top of standard classification and box-regression losses.

Experiments ran on a single NVIDIA RTX 4090. Classification used a Quaternion Wide ResNet in a 16-layer, width-4 configuration with AdamW, batch size 256, learning rate 0.1, and 300 epochs. Detection used YOLOv11n with SGD, batch size 32, automatic learning-rate scheduling, and 100 epochs.

Why This Matters

The paper argues that pure quaternion networks have been held back by implementation friction, not by theory: the absence of bounded non-constant analytic quaternion functions constrains activation design, and mainstream frameworks do not support quaternion arithmetic well. Showing that an approximation running on ordinary convolutions retains rotation awareness gives the field a practical path to rotation-equivariant models without specialized group-theoretic architectures.

Real-world applications:

• Robotic pick-and-place and assembly, where grasp planning depends on knowing object orientation, and where models must fit on constrained hardware.
• Aerial and remote sensing imagery, represented here by the DOTA-v1.0 benchmark with its oriented object annotations.
• 6D pose estimation and manipulation pipelines that need rotation-aware perception rather than axis-aligned boxes.
• Edge and embedded deployment in general, where the paper reports large parameter reductions relative to standard CNNs while keeping accuracy competitive.

Industry relevance comes from parameter counts in the sub-million to low-million range while retaining orientation awareness, which matters for real-time systems. The paper reports out-of-the-box improvements on a custom robotic dataset but degraded generalization on the larger COCO and DOTA benchmarks, so the practical value is currently strongest for narrow, domain-specific deployments rather than broad large-scale detection.

Future Directions

• Self-supervised learning with quaternion representations, extending the framework beyond supervised classification and detection.
• Optimizing the Hamilton mixing matrices themselves, rather than using the fixed sign pattern, to find task-specific structures.
• Adapting LoRA with quaternions for parameter-efficient fine-tuning of quaternion networks.
• Live deployment in robotic assembly, testing the framework in real manipulation settings where precise spatial understanding is required.

The paper also leaves open the question of whether the Poincaré mapping is optimal, noting it may not suit all downstream tasks, and whether the approximate convolution can be made to generalize better as dataset size grows, given the COCO and DOTA results.

Target Audience

Researchers and graduate students working on rotation-equivariant deep learning, quaternion neural networks, or oriented object detection, and robotics practitioners who need orientation-aware perception under tight compute and memory budgets. Readers should be comfortable with convolution arithmetic, backpropagation, and basic linear algebra; the quaternion algebra is developed from first principles in the methodology section, so prior familiarity with quaternions is helpful but not required. Engineers evaluating whether to replace a standard detector backbone with a parameter-efficient alternative will find the benchmark tables directly useful.

Authors’ abstract

This paper introduces Quaternion Approximate Networks (QUAN), a novel deep learning framework that leverages quaternion algebra for rotation equivariant image classification and object detection. Unlike conventional quaternion neural networks attempting to operate entirely in the quaternion domain, QUAN approximates quaternion convolution through Hamilton product decomposition using real-valued operations. This approach preserves geometric properties while enabling efficient implementation with custom CUDA kernels. We introduce Independent Quaternion Batch Normalization (IQBN) for training stability and extend quaternion operations to spatial attention mechanisms. QUAN is evaluated on image classification (CIFAR-10/100, ImageNet), object detection (COCO, DOTA), and robotic perception tasks. In classification tasks, QUAN achieves higher accuracy with fewer parameters and faster convergence compared to existing convolution and quaternion-based models. For objection detection, QUAN demonstrates improved parameter efficiency and rotation handling over standard Convolutional Neural Networks (CNNs) while establishing the SOTA for quaternion CNNs in this downstream task. These results highlight its potential for deployment in resource-constrained robotic systems requiring rotation-aware perception and application in other domains. Code: https://cwru-aism.github.io/ QUANpaper/