AF-SSD: An Accurate and Fast Single Shot Detector for High Spatial Remote Sensing Imagery

Ruihong Yin, Wei Zhao, Xudong Fan, Yongfeng Yin2020 · DOI

TL;DR

This paper introduces AF-SSD, a novel single-shot object detector designed for high-resolution remote sensing imagery. It addresses the challenges of detecting small, low-contrast, and few-texture objects by employing a lightweight backbone, an encoding-decoding module for small object enhancement, and a cascade attention mechanism for improved feature extraction, achieving state-of-the-art accuracy and speed.

What Problem Does This Paper Address?

Detecting objects in high-resolution remote sensing imagery presents unique challenges. Many existing object detection methods, particularly one-stage detectors like SSD and YOLO, struggle with accuracy when applied to this domain. This is primarily due to two factors:

  1. Small Object Detection: Remote sensing images often contain numerous small objects that lose significant detail in the down-sampling layers of deep convolutional neural networks (CNNs). High-level feature maps, crucial for semantic understanding, have limited resolution, making it difficult to detect these small targets.
  2. Low-Contrast and Few-Texture Objects: Objects like bridges, basketball courts, and tennis courts can have low contrast with their backgrounds or lack distinctive textures, making them hard for standard CNNs to differentiate from the background.

While some methods have improved accuracy, they often sacrifice speed, failing to meet the real-time detection requirements for applications like disaster prediction, search and rescue, and military surveillance. Therefore, there's a need for a detector that is simultaneously accurate, fast, and lightweight for these specific challenges.

Methods

The proposed AF-SSD builds upon the single-shot detector (SSD) architecture and introduces several novel components:

  • Lightweight Backbone: To reduce computational cost and memory footprint, the authors replace the VGG16 backbone of the original SSD with a custom lightweight architecture. This backbone utilizes depthwise separable convolutions (inspired by MobileNetV1) and incorporates multi-branch blocks from ShuffleNetV2. These modifications reduce parameters while maintaining or enhancing feature extraction capabilities.
  • Novel Encoding-Decoding Module (EDM): To address the small object detection problem, an EDM is integrated. This module fuses low-level, high-resolution features (rich in detail) with up-sampled high-level features (rich in semantics). This fusion enriches the semantic information available in low-level features, improving the detection of small objects.
  • Cascade Structure with Spatial and Channel Attention Modules: For detecting low-contrast and few-texture objects, a cascade attention mechanism is employed.
  • Spatial Attention Module: This module captures contextual information by calculating pixel-wise relationships within a feature map, helping to identify targets that lack internal distinctive features.
  • Channel Attention Module: This module adaptively weights feature channels, guiding the network to focus on the most informative features for distinguishing difficult targets. These two attention modules are cascaded sequentially to enhance feature discriminability.

  • Prediction Layers: The model uses four prediction layers on feature maps with strides {4, 8, 16, 32}, allowing it to detect objects at various scales.

Datasets

The primary dataset used for training and evaluation in this paper is the NWPU VHR-10 dataset. This dataset contains 800 very high-resolution remote sensing images with 10 object categories: Airplane, Ship, Storage tank, Baseball diamond, Tennis court, Basketball court, Ground track field, Harbor, Bridge, and Vehicle. The authors split this dataset into 20% for training, 20% for validation, and 60% for testing.

Key Insights and Ideas

  • Feature Fusion is Key for Small Objects: The paper strongly demonstrates that simply relying on high-level features is insufficient for small objects. The EDM's success in improving small object detection highlights the benefit of explicitly fusing detailed low-level features with semantic high-level features.
  • Attention for Context and Salience: The cascade attention module effectively tackles the challenges of low-contrast and few-texture objects. The spatial attention's ability to incorporate context and the channel attention's focus on salient features are crucial for these difficult-to-detect targets.
  • Lightweight Design Doesn't Mean Weak Performance: The authors show that a carefully designed lightweight backbone, incorporating modern efficient convolution techniques, can achieve comparable or better feature extraction than heavier backbones, while significantly reducing parameters and inference time.
  • Trade-off Management: The research explicitly aims to balance accuracy and speed, a common challenge in object detection. AF-SSD's performance suggests a successful strategy for achieving both in the context of remote sensing.

Key Results

The proposed AF-SSD achieved impressive results on the NWPU VHR-10 dataset:

  • Accuracy: An mAP of 88.7%, which is state-of-the-art compared to other methods evaluated on this dataset. It outperforms both two-stage and one-stage detectors.
  • Speed: An average running time of 0.035 seconds per image on an NVIDIA GTX-1080Ti GPU, making it a very fast detector.
  • Parameters: The model has only 5.7 million parameters, making it significantly more lightweight than many existing methods.
  • Ablation Studies: The paper demonstrates that each component contributes significantly:
  • The lightweight backbone alone achieved 80.8% mAP with only 4.3M parameters.
  • Adding the EDM boosted mAP to 86.5%, particularly improving small object detection.
  • The cascade attention module further improved performance to 88.2% mAP, especially for low-contrast and few-texture objects.
  • The final AF-SSD with Leaky ReLU achieved 88.7% mAP.

Limitations and Future Work

  • Limitations:
    • The paper focuses on a specific dataset (NWPU VHR-10). While it demonstrates strong performance, generalization to other remote sensing datasets with different characteristics might require further validation.
  • While the paper mentions the use of Leaky ReLU in the final model, the ablation study for Leaky ReLU shows a smaller improvement (0.5% mAP) compared to other modules, suggesting its contribution might be less critical than the EDM or attention modules.
  • Future Work:
    • The authors plan to apply their algorithms to embedded devices to test performance and inference time in real-world scenarios.
  • They also aim to research anchor-free detectors for more precise multi-scale object detection in remote sensing images.

Summary updated

BibTeX
@misc{yin2020af,
  author = {Ruihong Yin and Wei Zhao and Xudong Fan and Yongfeng Yin},
  title = {{AF-SSD: An Accurate and Fast Single Shot Detector for High Spatial Remote Sensing Imagery}},
  year = {2020},
  doi = {10.3390/s20226530},
}
Figure