PaperlayerView exact revision on arXiv

Source-linked AI summary

CFC-Net: A Critical Feature Capturing Network for Arbitrary-Oriented Object Detection in Remote Sensing Images

Qi Ming, Lingjuan Miao, Zhiqiang Zhou, Yunpeng Dong

arXiv:2101.06849v2cs.CV

TL;DR

Object detection in optical remote-sensing imagery is challenged by large variation in object scale, aspect ratio, and orientation. CFC-Net builds task-specific critical features, refines horizontal anchors, and dynamically selects anchors based on critical-feature capture ability. It reports superior performance across HRSC2016, DOTA, and UCAS-AOD, including best compared performance on the reported benchmarks.

  • Problem

    Remote-sensing object detection remains difficult because objects vary widely in scale, aspect ratio, and arbitrary orientation, limiting further performance improvement.

  • Method

    CFC-Net decouples classification and regression features, constructs task-specific critical features, refines horizontal anchors into rotation anchors, and dynamically selects high-quality anchors.

  • Results

    CFC-Net achieves superior detection performance across HRSC2016, DOTA, and UCAS-AOD, including mAP values of 86.3%, 73.50%, and 89.49% in reported settings.

  • Takeaways & Limitations

    The paper concludes that selecting anchors that capture critical features can support high-performance oriented detection with one preset anchor and reduced memory use.

Abstract

from arXiv · show

Object detection in optical remote sensing images is an important and challenging task. In recent years, the methods based on convolutional neural networks have made good progress. However, due to the large variation in object scale, aspect ratio, and arbitrary orientation, the detection performance is difficult to be further improved. In this paper, we discuss the role of discriminative features in object detection, and then propose a Critical Feature Capturing Network (CFC-Net) to improve detection accuracy from three aspects: building powerful feature representation, refining preset anchors, and optimizing label assignment. Specifically, we first decouple the classification and regression features, and then construct robust critical features adapted to the respective tasks through the Polarization Attention Module (PAM). With the extracted discriminative regression features, the Rotation Anchor Refinement Module (R-ARM) performs localization refinement on preset horizontal anchors to obtain superior rotation anchors. Next, the Dynamic Anchor Learning (DAL) strategy is given to adaptively select high-quality anchors based on their ability to capture critical features. The proposed framework creates more powerful semantic representations for objects in remote sensing images and achieves high-performance real-time object detection. Experimental results on three remote sensing datasets including HRSC2016, DOTA, and UCAS-AOD show that our method achieves superior detection performance compared with many state-of-the-art approaches. Code and models are available at https://github.com/ming71/CFC-Net.

I. INTRODUCTION

Remote-sensing object detection remains difficult because objects vary widely in scale, aspect ratio, and orientation, while preset horizontal anchors may include misleading background. CFC-Net addresses these challenges by constructing task-specific critical features, refining anchors, and selecting anchors that capture those features.

  • Background: CNN-based detectors use convolutional features and preset anchors for classification and regression, enabling efficient detection on large remote-sensing image collections.These frameworks have been applied to road, vehicle, airport, and ship detection.
  • Challenges: Remote-sensing objects’ varied scales, aspect ratios, and orientations make generic detectors difficult to adapt, especially when horizontal anchors contain substantial background around rotated targets.The background within horizontal boxes may mislead detection.
  • Challenges: Existing feature designs often target selected characteristics, while manually adding many anchors for scale, aspect-ratio, and orientation coverage increases anchor complexity without automatically capturing more complex discriminative features.The passage contrasts specialized feature representations with large manually specified anchor sets.
  • Critical features: Critical features are discriminative regions needed for accurate classification or localization, and they may be concentrated in local object areas rather than distributed evenly.Examples include an island, bow, or stern, and anchors that miss such regions can produce incorrect classifications.
  • Proposed approach: CFC-Net uses Polarization Attention to generate separate classification and regression feature pyramids, refines horizontal anchors into rotation anchors, and dynamically selects high-quality positives.The framework is designed around feature representation, anchor refinement, and training-sample selection.
  • Contributions: The paper reports that CFC-Net identifies critical features experimentally and uses them to optimize feature representation, anchor refinement, and training sample selection.Its stated contributions include interpreting common remote-sensing detection challenges from the critical-feature perspective.

II. RELATED WORK

The paper frames remote-sensing detection as difficult because objects vary greatly in scale, shape, and orientation, while classification and regression require different features. CFC-Net addresses these issues by decoupling task features, refining anchors, and dynamically selecting anchors that capture critical features.

  • Motivation: Remote-sensing objects vary widely in scale, aspect ratio, and orientation, making horizontal-anchor detection difficult and prone to background-induced errors.Horizontal boxes may contain substantial background when predicting rotated ground-truth boxes.
  • Anchor refinement and selection: CFC-Net refines preset horizontal anchors into rotation candidates using critical regression features, then dynamically selects anchors that capture those features for training.This design targets the alignment and classification-regression inconsistency problems in anchor-based detection.
  • CFC-Net: CFC-Net uses separate classification and regression feature pyramids with Polarization Attention to construct task-specific critical features.The classification branch emphasizes recognizable high-response features, whereas regression emphasizes boundaries and suppresses irrelevant high activations.
  • Polarization Attention Module: The Polarization Attention Module combines channel and spatial attention with the input feature pyramid to strengthen discriminative representations for detection.Channel attention models feature-map channel relationships, while spatial attention models dependencies between image pixels and uses dilated convolutions to expand receptive fields.

B. Rotation Anhcor Refinement Module

R-ARM generates rotation-aware anchor candidates from preset horizontal anchors using critical regression features. It reduces dependence on manually designed anchor geometries while avoiding classification-based filtering during refinement.

  • B. Rotation Anhcor Refinement Module: R-ARM regresses angle and box offsets from each preset horizontal anchor to produce a rotation anchor aligned with ground-truth objects.An initial anchor (x, y, w, h) becomes a rotated anchor (x, y, w, h, θ).
  • B. Rotation Anhcor Refinement Module: In CFC-Net, one initial anchor is preset at each feature-map position, eliminating the need to carefully specify anchor angle, aspect ratio, and scale.The paper attributes this reduced hyperparameter dependence to R-ARM's design after PAM.
  • B. Rotation Anhcor Refinement Module: Classification prediction is excluded from refinement because inaccurate early classification can remove high-quality candidates and classification confidence does not guarantee localization accuracy.The paper links classification-based sample selection to degraded detection performance under classification-regression mismatch.
  • B. Rotation Anhcor Refinement Module: Compared with previous one-stage detectors, CFC-Net uses fewer predefined anchors while achieving better detection performance with R-ARM.The framework's refined candidates align with high-response critical-feature regions.
  • B. Rotation Anhcor Refinement Module: R-ARM uses critical regression features to generate candidates whose regions capture discriminative object-boundary semantics for localization.The refined anchors also provide semantic priors for subsequent localization layers.

C. Dynamic Anchor Learning

Dynamic Anchor Learning addresses classification–localization misalignment by selecting anchors through matching degree and weighting training losses according to localization ability.

  • Dynamic Anchor Learning: DAL comprises dynamic anchor selection and matching-sensitive loss to select samples with strong critical-feature capturing ability during training.The strategy targets the gap between classification scores and localization quality.
  • Matching degree: Matching degree combines initial spatial alignment, post-regression alignment, and regression uncertainty to measure an anchor’s localization capacity.IoUin measures initial alignment, IoUout indicates critical regression-feature alignment, and u penalizes excessive IoU change.
  • Dynamic anchor selection: Anchors with matching degree above 0.6 are selected as positives, while each target without a positive receives its highest-matching candidate.Remaining anchors are treated as negatives.
  • Matching-sensitive loss: The weighting factor increases positive-sample contributions to classification loss and distinguishes candidates with different localization ability.This incorporates localization information into classification confidence, making selected detections more reliable.
  • Matching-sensitive loss: Matching-sensitive loss strengthens the correlation between classification score and localization ability compared with normal smooth-L1 training.The paper presents this as improving the reliability of detections selected by classification confidence.

A. Datasets

CFC-Net is evaluated on three public remote sensing datasets with oriented bounding-box annotations: HRSC2016, DOTA, and UCAS-AOD.

  • Datasets: Experiments use HRSC2016, DOTA, and UCAS-AOD, whose ground-truth boxes are annotated with oriented bounding boxes.These datasets cover ship, broad oriented-object, and aircraft-and-car detection settings.
  • HRSC2016: HRSC2016 contains 1,061 high-resolution remote sensing ship images divided into 436 training, 181 validation, and 444 test images.Images are resized to 416×416 and 800×800 in the experiments.
  • DOTA: DOTA contains 2,806 aerial images, 188,282 annotated instances, and 15 object categories for oriented object detection.Its categories include planes, vehicles, ships, bridges, and other aerial structures.
  • UCAS-AOD: UCAS-AOD contains 1,510 Google Earth images, including 1,000 plane images and 510 car images.The dataset is randomly split into training, validation, and test sets at a 5:2:3 ratio and resized to 800×800.

B. Implementation Details

Implementation uses a ResNet-50 backbone with multi-scale feature pyramids, one anchor per feature-map position, augmentation, and dataset-specific training schedules.

  • Model configuration: CFC-Net uses an ImageNet-pretrained ResNet-50 backbone fine-tuned on remote sensing datasets.Detection uses feature pyramid levels P3 through P7 for multi-scale objects.
  • Model configuration: Each feature-map position uses one anchor, while random flipping, rotation, and HSV jittering provide data augmentation.Positive matching-degree thresholds are 0.4 for refinement and 0.6 for detection.
  • Training: Training uses Adam with batch size 8 on an RTX 2080Ti, learning rate 1e-4, and 10-fold decay at each decay step.Total iterations are 10k for HRSC2016, 5k for UCAS-AOD, and 40k for DOTA.

1) Evaluation of different components:

Component-wise ablation on HRSC2016 shows that PAM, DAL, and R-ARM contribute cumulatively to CFC-Net’s detection performance.

  • Evaluation of different components:: The baseline reaches 70.5% mAP when only one anchor is preset, limiting its ability to capture critical object features.The experiment motivates components that improve feature representation and anchor quality.
  • Evaluation of different components:: DAL improves performance by 8.2% by selecting high-quality anchors with better critical-feature alignment during learning.The reported gain is attributed to dynamic selection of anchors suited to localization.
  • Evaluation of different components:: Using DAL and PAM together produces 83.8% mAP, indicating that their effects are compatible in the ablation.The two methods are reported not to conflict.
  • Evaluation of different components:: R-ARM further improves performance by 2.5% by refining horizontal anchors into high-quality rotated anchors.This extends the gains from feature construction and dynamic anchor selection.
  • Evaluation of different components:: 86.3% mAP is achieved by CFC-Net, a 15.8% increase over the naive model.The result is reported as evidence for the effectiveness of the complete framework.

2) Evaluation of PAM:

PAM improves detection by separating classification and regression features, then emphasizing task-specific critical regions. Ablations show gains from dual FPN, attention, and polarization-based feature processing.

  • Dual FPN improves detection performance by 1.6% over the baseline by extracting independent classification and regression features.
  • Adding attention to dual FPN yields a further 2.8% improvement by helping each branch respond to discriminative object parts.
  • Polarization processing strengthens discriminative classification regions and suppresses high-response regression regions to expose additional localization clues.
  • The polarization-based attention model improves performance by 1.3% over the attention-based model, supporting PAM’s effectiveness.
  • PAM heatmaps emphasize local classification regions, such as ship sterns and bows, while regression cues concentrate on object edges and context.

4) Evaluation of R-ARM:

R-ARM and related anchor-learning choices improve localization and benchmark performance while reducing reliance on many preset anchors. The experiments also show that hyperparameter balance matters for stable anchor selection.

  • Evaluation of R-ARM: With one-stage R-ARM, refinement and detection thresholds are set to 0.4 and 0.6, respectively, within the DAL-and-PAM framework.
  • Evaluation of R-ARM: As α decreases from 0.9 to 0.5 with γ = 4, mAP increases from 72.1% to 78.7%.
  • Evaluation of R-ARM: When α becomes extremely small, dominant output IoU can select too many false positives and sharply reduce performance.
  • Results on HRSC2016: CFC-Net reaches 86.3% mAP at 416×416 without augmentation, 88.6% with augmentation at 800×800, and 89.7% with multiscale training and testing on HRSC2016.
  • Results on HRSC2016: Using one horizontal anchor per feature-map position, the method outperforms frameworks with many anchors and reaches 28 FPS at 800×800 on an RTX 2080 Ti.

2) Results on DOTA:

CFC-Net performs strongly on DOTA and related remote-sensing benchmarks, including densely arranged, differently scaled, and highly elongated objects. The reported results support using critical-feature capture rather than extensive preset-anchor alignment.

  • Results on DOTA: CFC-Net achieves 73.50% mAP on DOTA, the best performance among the compared methods.
  • Results on DOTA: With only one anchor, the method detects densely arranged small objects and adapts to substantial scale variation.
  • Results on DOTA: Square anchors can detect objects with very large aspect ratios, including bridges and harbors, when critical features are effectively captured.
  • Results on UCAS-AOD: On another benchmark, CFC-Net reaches 89.49% mAP and outperforms the compared single-stage detectors and some advanced two-stage detectors.
  • Conclusion: The conclusion attributes the framework’s high-quality detection to critical-feature extraction, rotation-anchor refinement, and matching-degree-based positive selection.
Loading 2101.06849v2…