Sparse Label Assignment for Oriented Object Detection in Aerial Images

The mainstream algorithms used for ship classification and detection can be improved based on convolutional neural networks (CNNs). By analyzing the characteristics of ship images, we found that the difficulty in ship image classification lies in distinguishing ships with similar hull structures but different equipment and superstructures. To extract features such as ship superstructures, this paper introduces transformer architecture with self-attention into ship classification and detection, and a CNN and Swin transformer model (CNN-Swin model) is proposed for ship image classification and detection. The main contributions of this study are as follows: (1) The proposed approach pays attention to different scale features in ship image classification and detection, introduces a transformer architecture with self-attention into ship classification and detection for the first time, and uses a parallel network of a CNN and a transformer to extract features of images. (2) To exploit the CNN’s performance and avoid overfitting as much as possible, a multi-branch CNN-Block is designed and used to construct a CNN backbone with simplicity and accessibility to extract features. (3) The performance of the CNN-Swin model is validated on the open FGSC-23 dataset and a dataset containing typical military ship categories based on open-source images. The results show that the model achieved accuracies of 90.9% and 91.9% for the FGSC-23 dataset and the military ship dataset, respectively, outperforming the existing nine state-of-the-art approaches. (4) The good extraction effect on the ship features of the CNN-Swin model is validated as the backbone of the three state-of-the-art detection methods on the open datasets HRSC2016 and FAIR1M. The results show the great potential of the CNN-Swin backbone with self-attention in ship detection.

Section: Performance Comparison For Ship Detectionmentioning

confidence: 99%

Section: Performance Comparison For Ship Detectionmentioning

confidence: 99%

Fine-Grained Ship Classification by Combining CNN and Swin Transformer

Huang

Zhang

2022

“…Currently, mainstream oriented object detectors [20][21][22][23] are based on densely placed predefined anchors. Several early rotation detectors use a horizontal anchor-based Region Proposal Network (RPN) to generate horizontal regions of interest (RoIs), and then design novel network modules to convert the horizontal RoIs into OBBs.…”

Section: Introductionmentioning

confidence: 99%

Oriented Object Detection in Remote Sensing Images with Anchor-Free Oriented Region Proposal Network

Tian

et al. 2022

Oriented object detection is a fundamental and challenging task in remote sensing image analysis that has recently drawn much attention. Currently, mainstream oriented object detectors are based on densely placed predefined anchors. However, the high number of anchors aggravates the positive and negative sample imbalance problem, which may lead to duplicate detections or missed detections. To address the problem, this paper proposes a novel anchor-free two-stage oriented object detector. We propose the Anchor-Free Oriented Region Proposal Network (AFO-RPN) to generate high-quality oriented proposals without enormous predefined anchors. To deal with rotation problems, we also propose a new representation of an oriented box based on a polar coordinate system. To solve the severe appearance ambiguity problems faced by anchor-free methods, we use a Criss-Cross Attention Feature Pyramid Network (CCA-FPN) to exploit the contextual information of each pixel and its neighbors in order to enhance the feature representation. Extensive experiments on three public remote sensing benchmarks—DOTA, DIOR-R, and HRSC2016—demonstrate that our method can achieve very promising detection performance, with a mean average precision (mAP) of 80.68%, 67.15%, and 90.45%, respectively, on the benchmarks.

“…Therefore, the research of rotated object detection in remote sensing images is of great significance for engineering applications. In recent years, rotated object detection has been derived from classic object detection [27][28][29], and most existing methods use five parameters (coordinates of the central point, width, height, and rotation angle) to describe the oriented bounding box. The initial exploration of rotated object detection involves rotating the RPN [30]; however, it involves more anchors, which implies that additional running time is required.…”

Section: Introductionmentioning

confidence: 99%

Constraint Loss for Rotated Object Detection in Remote Sensing Images

Zhang

Wang

et al. 2021

Rotated object detection is an extension of object detection that uses an oriented bounding box instead of a general horizontal bounding box to define the object position. It is widely used in remote sensing images, scene text, and license plate recognition. The existing rotated object detection methods usually add an angle prediction channel in the bounding box prediction branch, and smooth L1 loss is used as the regression loss function. However, we argue that smooth L1 loss causes a sudden change in loss and slow convergence due to the angle solving mechanism of openCV (the angle between the horizontal line and the first side of the bounding box in the counter-clockwise direction is defined as the rotation angle), and this problem exists in most existing regression loss functions. To solve the above problems, we propose a decoupling modulation mechanism to overcome the problem of sudden changes in loss. On this basis, we also proposed a constraint mechanism, the purpose of which is to accelerate the convergence of the network and ensure optimization toward the ideal direction. In addition, the proposed decoupling modulation mechanism and constraint mechanism can be integrated into the popular regression loss function individually or together, which further improves the performance of the model and makes the model converge faster. The experimental results show that our method achieves 75.2% performance on the aerial image dataset DOTA (OBB task), and saves more than 30% of computing resources. The method also achieves a state-of-the-art performance in HRSC2016, and saved more than 40% of computing resources, which confirms the applicability of the approach.