Source-linked AI summary
Poly Kernel Inception Network for Remote Sensing Detection
Xinhao Cai, Qiuxia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, Yazhou Yao
TL;DR
Remote sensing detection must handle large object-scale variation and diverse context, while large-kernel and dilated convolutions introduce background noise or sparse representations. PKINet combines parallel multi-scale depth-wise convolutions without dilation with Context Anchor Attention, achieving state-of-the-art performance on four remote sensing benchmarks.
Problem
Remote sensing images contain objects with widely varying scales and recognition depends on diverse surrounding context, while existing receptive-field expansion can introduce background noise or sparse features.
Method
PKINet uses parallel depth-wise convolution kernels of different sizes without dilation for dense multi-scale texture features, together with Context Anchor Attention for long-range context.
Results
PKINet achieves state-of-the-art performance on four remote sensing benchmark datasets: DOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R.
Takeaways & Limitations
The proposed backbone addresses scale variation and contextual diversity in remote sensing object detection while remaining lightweight through depth-wise and 1D convolutions.
Takeaways & Limitations
Limited computational resources prevented scaling PKINet's model capacity to assess its maximal potential, leaving scalability for future work.
Abstract
from arXiv · showhide
Object detection in remote sensing images (RSIs) often suffers from several increasing challenges, including the large variation in object scales and the diverse-ranging context. Prior methods tried to address these challenges by expanding the spatial receptive field of the backbone, either through large-kernel convolution or dilated convolution. However, the former typically introduces considerable background noise, while the latter risks generating overly sparse feature representations. In this paper, we introduce the Poly Kernel Inception Network (PKINet) to handle the above challenges. PKINet employs multi-scale convolution kernels without dilation to extract object features of varying scales and capture local context. In addition, a Context Anchor Attention (CAA) module is introduced in parallel to capture long-range contextual information. These two components work jointly to advance the performance of PKINet on four challenging remote sensing detection benchmarks, namely DOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R.
1. Introduction
Remote sensing detection must handle objects spanning widely different scales and use surrounding context, but receptive-field expansion can trade fine detail for noise. PKINet combines multi-scale convolution and contextual attention to address these challenges with a lightweight backbone.
- RSI objects range from soccer fields to vehicles, and accurate recognition depends on both appearance and surrounding environmental context.
- Large-kernel convolutions can introduce background noise for small objects, while dilated convolutions may overlook fine-grained details and produce sparse features.
- PKINet uses parallel depth-wise kernels of different sizes without dilation to extract dense multi-scale texture features and fuse local contextual information.
- A Context Anchor Attention mechanism uses global average pooling and 1D strip convolutions to capture long-range relationships between distant pixels.
- PKINet explores inception-style convolutions and global context attention for remote sensing detection while remaining lightweight through depth-wise and 1D convolutions.
2. Related Work
Prior remote sensing detectors often rely on standard backbones that overlook RSI-specific scale variation and contextual diversity. PKINet instead combines inception-style depth-wise multi-scale extraction with Context Anchor Attention.
- Remote sensing detection faces arbitrary object orientations, substantial scale variations, and diverse contextual information.
- OBB for Remote Sensing Object Detection: Specialized oriented bounding-box detectors improve orientation handling but often retain standard backbones that overlook RSI-specific characteristics.
- Feature Extraction for Remote Sensing Object Detection: Existing feature-extraction methods address scale variation through data augmentation, multi-scale integration, or feature-pyramid hierarchies.
- PKINet replaces large-kernel or dilated receptive-field expansion with inception-style depth-wise convolutions without dilation for multi-scale texture features.
- Its Context Anchor Attention mechanism captures long-range contextual information, complementing local multi-scale extraction.
3. Methodology
PKINet combines a four-stage CSP-style backbone with inception-style multi-scale feature extraction and Context Anchor Attention. Its PKI and CAA modules jointly capture local, multi-scale, and long-range contextual information without dilated convolution.
- PKI Stage: PKINet uses four stages with CSP structures that split each stage input into an FFN path and a PKI Block path before concatenation.The stage output is produced by concatenating both paths and applying a 1×1 convolution.
- PKI Module: The PKI Module applies a small-kernel convolution followed by parallel depth-wise convolutions with different kernel sizes to extract multi-scale local and contextual features.The module uses ks = 3 and k(m) = (m+1)×2+1 in the reported configuration.
- PKI Module: The PKI Module does not use dilated convolution, thereby avoiding overly sparse feature representations.
- PKI Module: A 1×1 convolution fuses features from different receptive-field sizes, integrating broad contextual information while preserving local texture features.The fused feature is formed from the local feature and four contextual features.
- Context Anchor Attention (CAA): CAA captures long-range contextual information using average pooling, 1×1 convolution, and depth-wise strip convolutions within each PKI Block.Strip convolutions approximate a standard large-kernel depth-wise convolution and support slender-shape feature extraction; kb is set as 11+2×l.
- Context Anchor Attention (CAA): The CAA attention map is constrained to (0, 1), then element-wise multiplied with and added to the PKI output to produce an enhanced feature.A final 1×1 convolution produces the PKI Block output.
4. Experiment
PKINet is evaluated across remote sensing benchmarks, detector architectures, visual comparisons, and targeted ablations. Results show improved detection performance, efficiency, scale robustness, and benefits from its multi-scale kernels, CAA, and CSP design.
- DOTA-v1.0: PKINet-T improves over ResNet-18 by 3.67% using 36.7% of its parameters and 59.6% of its compute on DOTA-v1.0.PKINet-S improves over ResNet-50 by 2.52% with 58.8% of the parameters and 81.53% of the compute.
- DOTA-v1.0: PKINet improves mAP over ResNet-50 by 2.41%, 6.19%, and 3.71% with Rotated FCOS, R3Det, and S2ANet, respectively.With Oriented RCNN, performance reaches 78.39%, including gains of 5.3% and 5.76% for SV and LV over LSKNet, and 6.46% for RA.
- Cross-dataset results: PKINet achieves a 1.21% improvement on DOTA-v1.5, surpasses 12 leading methods on HRSC2016, and reaches 67.03% on DIOR-R.The DOTA-v1.5 result is reported on a more challenging dataset containing many minuscule instances.
- Visual analysis: Visual results indicate that PKINet detects large targets while retaining focus on smaller targets across scenes with substantial size variation.The comparison is made against LSKNet on DOTA.
- Ablation studies: A 3×3-to-11×11 multi-scale kernel design performs best, while larger-only kernels decrease performance by 0.49% and 0.84%.Using only 3×3 kernels provides limited texture information, whereas stride 4 is sub-optimal relative to the adopted stride 2 design.
- Ablation studies: Five kernels yield optimal performance, dilation reduces performance by 1.09%, CAA across all stages gains 1.03%, and removing CSP raises parameters and computation by 211% and 159%.The ablations link kernel count, dilation, CAA placement, and CSP structure to performance or efficiency changes.
5. Discussion and Conclusion
PKINet combines parallel multi-scale depth-wise convolutions with Context Anchor Attention to address object-scale variation and contextual diversity in remote sensing images. It achieves state-of-the-art performance on four benchmark datasets, while scalability beyond the tested model capacities remains future work.
- PKINet uses parallel depth-wise convolution kernels of various sizes to capture dense texture features across different scales.A Context Anchor Attention mechanism further captures long-range contextual information.
- PKINet achieves state-of-the-art performance on four remote sensing benchmark datasets.
- Limited computational resources prevented scaling PKINet capacity to its maximal potential, leaving model scalability for future research.