Advances in synthetic aperture radar (SAR) technology, combined with innovations in deep learning algorithms, have established SAR as a critical method for ship detection. However, in some complex environments, such as harbors and inshore areas, false and missed detections tend to occur because of intricate background interference and the minute proportions of ship targets. To overcome these issues, this paper presents a ship detection model, CSS-YOLO, for SAR images in complex scenes. First, a vertically compressed backbone network (VCCB) is proposed, which reduces feature loss and enhances the extraction of detailed features by decreasing the number of downsampling and increasing the feature extraction width. Second, a wavelet convolution-based preprocessing downsampling module (WPCM) is proposed to obtain a larger receptive field and reduce noise to obtain finer features. Third, a short-range aggregated decentralized necking network (SADNet) is proposed to reduce the loss of detail, thereby improving the efficiency of feature fusion. To validate the effectiveness of the method, this paper conducts experiments on three datasets, HRSID, SSDD, and LS-SSDD-v1.0, for performance evaluation. Compared with other widely used methods, CSS-YOLO demonstrates excellent performance with mAP0.5 values of 94.8%, 98.6%, and 78.4% on HRSID, SSDD, and LS-SSDD-v1.0 datasets, respectively. Evaluation results demonstrate that CSS-YOLO outperforms state-of-the-art SAR ship detection approaches in accuracy and overall performance.

Unmanned aerial vehicle (UAV) image object detection has become an increasingly important research area in computer vision. However, the variable target shapes and complex environments make it difficult for the model to fully exploit its features. In order to solve this problem, we propose a UAV image object detection method based on a backbone feature reuse detection network, named BFRDNet. First, we design a backbone feature reuse pyramid network (BFRPN), which takes the model characteristics as the starting point and more fully utilizes the multi-scale features of backbone network to improve the model’s performance in complex environments. Second, we propose a feature extraction module based on multiple kernels convolution (MKConv), to deeply mine features under different receptive fields, helping the model accurately recognize targets of different sizes and shapes. Finally, we design a detection head preprocessing module (PDetect) to enhance the feature representation fed to the detection head and effectively suppress the interference of background information. In this study, we validate the performance of BFRDNet primarily on the VisDrone dataset. The experimental results demonstrate that BFRDNet achieves a significant improvement in detection performance, with the mAP increasing by 7.5%. To additionally evaluate the model’s generalization capacity, we extend the experiments to the UAVDT and COCO datasets.

Infrared images usually lack significant texture and morphological features, while there is a large amount of noise and clutter in the background. These characteristics make it difficult for general-purpose target detection algorithms to adequately extract key features of small targets in infrared images, resulting in incomplete feature information. In addition, the direct feature fusion of traditional neck networks suffers from feature misalignment, and the features of small targets are easily diluted or masked, further exacerbating the problem of feature loss. To solve the above-mentioned problems, this article proposes a dual-stream cooperative sensing detection network (DCSDNet) specifically designed for infrared small target detection. First, to improve the feature extraction capability of the model, this article proposes the dense grouped feature learning module, which provides a richer feature representation base for the infrared small target detection task through multilevel feature processing. Meanwhile, to reduce the problems of information loss and fusion imbalance, the refined feature fusion bottleneck module is designed to achieve a finer feature fusion, which improves the feature extraction accuracy and fusion effect. In addition, this article critically re-examines the necessity of feature pyramids through experimental validation and heatmap visualization. We deliberately discard the conventional pyramid structure prevalent in multiobject detection systems, and propose a feature harmonization backbone. This innovation not only reduces feature dilution and misalignment risks but also strengthens the model’s feature extraction capacity. Ultimately, combining these improvements forms the DCSDNet target detection algorithm, which focuses on improving the detection of small infrared targets. The DCSDNet method proposed in this article achieves a mAP0.5 of 68.1%, 95.4%, and 83.4% on the VEDAI-Infrared, NUAA-SIRST, and IRSTD-1 K datasets, respectively. These data indicate that the algorithm achieves deep feature extraction of small targets in infrared images and shows superior performance in IR small target detection.