Class imbalance problem exists widely in vision data. In these imbalanced datasets, the majority classes dominate the loss and influence the gradient. Hence, these datasets have a significantly negative impact on the performance of many state of the art methods. In this paper, we propose a class imbalance loss (CI loss) to handle this problem. To distinguish imbalanced datasets in accordance with the extent of imbalance, we also define an imbalance degree that works as a decision index factor in the CI loss. Because the minority classes with fewer samples probably lose chances in descending gradient in the training process, CI loss is introduced to make these minority classes descend further than the majority classes. In view of the imbalanced distribution of data in few-shot learning, a method for generating an imbalanced few-shot learning dataset is presented in this paper. We conducted a large number of experiments in the MiniImageNet dataset, which showed the effectiveness of an algorithm for model-agnostic meta-learning (MAML) for rapid adaptation with CI loss. In the problem of detecting 15 ship categories, our loss function is transplanted to Rotational Region Convolutional Neural Network (R2CNN) detection method and Cascade network architecture and achieves higher mAP than focal loss and cross-entropy loss. In addition, Mixed National Institute of Standards and Technology dataset and Moving and Stationary Target Acquisition and Recognition dataset are sampled to imbalanced datasets to verify the effectiveness of CI loss.

Oriented object detection in remote sensing image interpretation is challenging due to the difficulty of locating objects with arbitrary orientations. Existing methods have made considerable progress based on oriented heads or anchors. However, most of them follow the classical detection paradigm, such as assigning samples based on Intersection-over-Unions (IoU) and predicting through two independent tasks. These fixed strategies impair the consistency between classification and localization predictions, resulting in the prediction with optimal localization accuracy being suppressed by the non-optimal ones during Non-Maximum Suppression (NMS). In order to address this problem, a Task-Collaborated Detector (TCD) is proposed. Compared with current single-stage methods, its improvements include two aspects: Task-Collaborated Assignment (TCA) and Task-Collaborated Head (TCH). Specifically, in order to better pull closer the best anchors for two tasks, TCA introduces classification and localization confidence into sample assignment and tends to select the anchors with accurate and consistent predictions as positive during training. TCH provides a better balance for learning interactive and discriminative features. It can flexibly adjust the spatial feature distribution of classification and localization tasks by learning the joint features from the aggregation layer. Extensive experiments are conducted on HRSC2016, DOTA, and DIOR-R, and the proposed TCD achieves state-of-the-art performance (90.60, 80.89, and 65.04 mAP, respectively). Consistency analysis also demonstrates that TCD can significantly improve prediction consistency.

Although existing oriented object detection methods have made considerable progress based on oriented heads or anchors, the training process itself is not perfect. In this letter, we point out the inconsistency problem between the fixed network setting and varying aspect ratios, which greatly limits the performance. For example, the fixed parameters in label assignment and regression loss cannot fit the changes of aspect ratios and, thus, are harmful to the training process. Considering the prior information about objects’ aspect ratios, the aspect-ratio-guided (ARG) methods are proposed. Specifically, the ARG label assignment is used to adjust the label assignment criteria (intersection over union (IoU) threshold) automatically, and the ARG IoU loss can change the weights of angle regression dynamically. This ARG design makes better use of training samples and pushes the detector more robust to the change of aspect ratios. With no additional cost, our method improves upon the ResNet-50-feature pyramid network (FPN) baseline with 3.99% AP50 and 6.09% AP75 on HRSC2016.

Multioriented object detection, an important yet challenging task because of the bird's-eye-view perspective, complex background, and densely packed objects, is in the spotlight of detection in remote sensing images. Although existing methods have recently experienced substantial progress based on oriented head, they learn little about essential rotation invariance of the objects. In this article, a novel framework is proposed that can learn high-quality rotation invariance features of the multioriented objects by three measures. Given a remote sensing image, the multiscale semantic segmentation feature fusion module first merges the global semantic segmentation features predicted by the semantic segmentation branch and the multiscale features extracted by the backbone with FPN in order to distinguish complex background. Then, the discriminative features are used by rotation mainstream, whose structure is similar to cascade R-CNN and can extract higher quality rotation invariance features and predict more accurate location information by adaptively adjusting the distribution of the samples through progressive intersection over union thresholds. And in order to improve the performance of mainstream to predict more accurate oriented bounding box, the horizontal tributaries that can fully leverage the reciprocal relationship between the oriented detection and horizontal detection were added to the latter two stages. Extensive experiments on three public datasets for remote sensing images, i.e., Gaofen Airplane, HRSC2016, and DOTA demonstrate that without bells and whistles, the proposed method achieves superior performances compared with the existing state-of-the-art methods for multioriented detection. Moreover, our overall system achieves 59.264% mAP of airplane Detection in 2020 Gaofen challenge, ranking third in the final.

Synthetic Aperture Radar image object detection holds significant application value in both military and civilian domains. However, existing deep learning-based methods suffer from excessive model parameters and high computational costs, making them impractical for real-time deployment on edge computing platforms. To address these challenges, this paper proposes a lightweight SAR object detection method optimized for edge devices. First, we design an efficient backbone network based on inverted residual blocks and the information bottleneck principle, achieving an optimal balance between feature extraction capability and computational resource consumption. Then, a Fast Feature Pyramid Network is constructed to enable efficient multi-scale feature fusion. Finally, we propose a decoupled network-in-network Head, which significantly reduces the computational overhead while maintaining detection accuracy. Experimental results demonstrate that the proposed method achieves comparable detection performance to state-of-the-art YOLO variants while drastically reducing computational complexity (4.4 GFLOP) and parameter count (1.9 M). On edge platforms (Jetson TX2 and Huawei Atlas DK 310), the model achieves real-time inference speeds of 34.2 FPS and 30.7 FPS, respectively, proving its suitability for resource-constrained, real-time SAR object detection scenarios.