Although remarkable progress has been made in salient object detection (SOD) in natural scene images (NSI), the SOD of optical remote sensing images (RSI) still faces significant challenges due to various spatial resolutions, cluttered backgrounds, and complex imaging conditions, mainly for two reasons: (1) accurate location of salient objects; and (2) subtle boundaries of salient objects. This paper explores the inherent properties of multi-level features to develop a novel semantic-guided attention refinement network (SARNet) for SOD of NSI. Specifically, the proposed semantic guided decoder (SGD) roughly but accurately locates the multi-scale object by aggregating multiple high-level features, and then this global semantic information guides the integration of subsequent features in a step-by-step feedback manner to make full use of deep multi-level features. Simultaneously, the proposed parallel attention fusion (PAF) module combines cross-level features and semantic-guided information to refine the object’s boundary and highlight the entire object area gradually. Finally, the proposed network architecture is trained through an end-to-end fully supervised model. Quantitative and qualitative evaluations on two public RSI datasets and additional NSI datasets across five metrics show that our SARNet is superior to 14 state-of-the-art (SOTA) methods without any post-processing.

Drone-based object detection has been widely applied in ground object surveillance, urban patrol, and some other fields. However, the dramatic scale changes and complex backgrounds of drone images usually result in weak feature representation of small objects, which makes it challenging to achieve high-precision object detection. Aiming to improve small objects detection, this paper proposes a novel cross-scale knowledge distillation (CSKD) method, which enhances the features of small objects in a manner similar to image enlargement, so it is termed as ZoomInNet. First, based on an efficient feature pyramid network structure, the teacher and student network are trained with images in different scales to introduce the cross-scale feature. Then, the proposed layer adaption (LA) and feature level alignment (FA) mechanisms are applied to align the feature size of the two models. After that, the adaptive key distillation point (AKDP) algorithm is used to get the crucial positions in feature maps that need knowledge distillation. Finally, the position-aware L2 loss is used to measure the difference between feature maps from cross-scale models, realizing the cross-scale information compression in a single model. Experiments on the challenging Visdrone2018 dataset show that the proposed method draws on the advantages of the image pyramid methods, while avoids the large calculation of them and significantly improves the detection accuracy of small objects. Simultaneously, the comparison with mainstream methods proves that our method has the best performance in small object detection.

In real-life complex traffic environments, vehicles are often occluded by extraneous background objects and other vehicles, leading to severe degradation of object detector performance. To address this issue, we propose a method named YOLO-OVD (YOLO for occluded vehicle detection) and a dataset for effectively handling vehicle occlusion in various scenarios. To highlight the model attention in unobstructed region of vehicles, we design a novel grouped orthogonal attention (GOA) module to achieve maximum information extraction between channels. We utilize grouping and channel shuffling to address the initialization and computational issues of original orthogonal filters, followed by spatial attention for enhancing spatial features in vehicle-visible regions. We introduce a CIoU-based repulsion term into the loss function to augment the network’s localization accuracy in scenarios involving densely packed vehicles. Moreover, we explore the effect of the knowledge-based Laplacian Pyramid on the OVD performance, which contributes to fast convergence in training and ensures more detailed and comprehensive feature retention. We conduct extensive experiments on the established occluded vehicle detection dataset, which demonstrates that the proposed YOLO-OVD model significantly outperforms 14 representative object detectors. Notably, it achieves improvements of 4.7% in Precision, 3.6% in AP@0.5, and 1.9% in AP@0.5:0.95 compared to the YOLOv5 baseline.

In recent years, numerous advanced lightweight models have been proposed for salient object detection (SOD) in optical remote sensing images (ORSI). However, most methods still face challenges such as performance limitations and imbalances between accuracy and computational cost. To address these issues, we propose SggNet, a novel semantic- and graph-guided lightweight network for ORSI-SOD. The SggNet adopts a classical encoder-decoder structure with MobileNet-V2 as the backbone, ensuring optimal parameter utilization. Furthermore, we design an Efficient Global Perception Module (EGPM) to capture global feature relationships and semantic cues through limited computational costs, enhancing the model’s ability to perceive salient objects in complex scenarios, and a Semantic-Guided Edge Awareness Module (SEAM) that leverages the semantic consistency of deep features to suppress background noise in shallow features, accurately predict object boundaries, and preserve the detailed shapes of salient objects. To further efficiently aggregate multi-level features and preserve the integrity and complexity of overall object shape, we introduce a Graph-Based Region Awareness Module (GRAM). This module incorporates non-local operations under graph convolution domain to deeply explore high-order relationships between adjacent layers, while utilizing depth-wise separable convolution blocks to significantly reduce computational cost. Extensive quantitative and qualitative experiments demonstrate that the proposed model achieves excellent performance with only 2.70 M parameters and 1.38 G FLOPs, while delivering an impressive inference speed of 108 FPS, striking a balance between efficiency and accuracy to meet practical application needs.

The targets in Synthetic Aperture Radar (SAR) images are often tiny, irregular, and difficult to detect against complex backgrounds, leading to a high probability of missed or incorrect detections by object detection algorithms. To address this issue and improve the recall rate, we introduce an improved version of YOLOv8 (You Only Look Once), named MDD-YOLOv8. This model is not only fast but also highly accurate, with fewer instances of missed or incorrect detections. Our proposed model outperforms the baseline YOLOv8 in SAR image detection by utilizing dynamic convolution to replace static convolution (DynamicConv) and incorporating a deformable large kernel attention mechanism (DLKA). Additionally, we modify the structure of the FPN-PAN and introduce an extra detection header to better detect tiny objects. Experiments on the MSAR-1.0 dataset demonstrate that MDD-YOLOv8 achieves 87.7% precision, 76.1% recall, 78.9% mAP@50, and 0.81 F1 score. These metrics show an improvement of 8.1%, 6.0%, 6.9%, and 0.07, respectively, compared to the original YOLOv8. Although, MDD-YOLOv8 increases parameters by about 20% and GFLOPs by 53% more than YOLOv8n. To further validate the model’s effectiveness, we conducted generalization experiments on four additional SAR image datasets, proving that MDD-YOLOv8’s performance is robust and universally applicable. In summary, MDD-YOLOv8 is a robust, generalized model with strong potential for industrial applications.

The majority of modern object detectors rely on a set of pre-defined anchor boxes, which enhances detection performance dramatically. Nevertheless, the pre-defined anchor strategy suffers some drawbacks, especially the complex hyper-parameters of anchors, seriously affecting detection performance. In this paper, we propose a feature-guided anchor generation method named dynamic anchor. Dynamic anchor mainly includes two structures: the anchor generator and the feature enhancement module. The anchor generator leverages semantic features to predict optimized anchor shapes at the locations where the objects are likely to exist in the feature maps; by converting the predicted shape maps into location offsets, the feature enhancement module uses the high-quality anchors to improve detection performance. Compared with the hand-designed anchor scheme, dynamic anchor discards all pre-defined boxes and avoids complex hyper-parameters. In addition, only one anchor box is predicted for each location, which dramatically reduces calculation. With ResNet-50 and ResNet-101 as the backbone of the one-stage detector RetinaNet, dynamic anchor achieved 2.1 AP and 1.0 AP gains, respectively. The proposed dynamic anchor strategy can be easily integrated into the anchor-based detectors to replace the traditional pre-defined anchor scheme.