To address the problems of missed and false detections caused by insufficient feature quality in small object detection from UAV perspectives, this paper proposes a UAV small object detection algorithm based on YOLOv8 feature optimization. A spatial cosine convolution module is introduced into the backbone network to optimize spatial features, thereby alleviating the problem of small object feature loss and improving the detection accuracy and speed of the model. An improved C2f_SCConv feature fusion module is employed for feature integration, which effectively reduces feature redundancy in spatial and channel dimensions, thereby lowering model complexity and computational cost. Meanwhile, the WIoU loss function is used to replace the original CIoU loss function, reducing the interference of geometric factors in anchor box regression, enabling the model to focus more on low-quality anchor boxes, and enhancing its small object detection capability. Ablation and comparative experiments on the VisDrone dataset validate the effectiveness of the proposed algorithm for small object detection from UAV perspectives, while generalization experiments on the DOTA and SSDD datasets demonstrate that the algorithm possesses strong generalization performance.

Since deep learning has been applied to medical image analysis, convolutional neural networks (CNNs) and Vision Transformers (ViTs) have become core architectures for medical image classification tasks. However, CNNs are constrained by local receptive fields and cannot effectively model long-range context-dependent correlations in medical images; ViTs face deployment challenges in high-resolution medical image tasks due to the quadratic computational complexity of self-attention. While State Space Models (SSMs) represented by Mamba offer a new solution with linear complexity, they suffer from redundant directional modeling and low parameter efficiency in direct medical image applications. This study proposes the efficient state space enhancement framework (MedMamba-ESS), integrating the SS2D-Top2 adaptive directional scanning mechanism (reducing SSM submodule FLOPs by ~37%) and the G-SSM grouped parameter sharing module (achieving 3–4% parameter compression and 4.4% accuracy improvement via regularization). Validated on 14 public datasets (9 imaging modalities, 9 anatomical regions, >240,000 images), MedMamba-ESS achieves superior performance at 2.0G FLOPs: its Overall Accuracy (OA) is 3% higher than the baseline MedMamba-Tiny on non-MedMNIST datasets (3.13% higher than other mainstream models) and 3% higher on MedMNIST datasets (1% higher than others). Ablation experiments confirm that the two modules synergistically reduce parameters by 1.05% and boost accuracy by 4.6%. This study overcomes the technical limitations of traditional SSMs in medical imaging applications, achieving synergistic improvements in both model performance and computational efficiency. It provides an architecture optimization solution that combines practicality and generalizability for the implementation of SSMs in medical image analysis.

High-resolution remote sensing image change detection holds significant application value in the fields of urban planning, disaster assessment, and others. However, it faces the dual challenge of pseudo-change interference and loss of detailed information. To address these issues, a frequency-domain-aware Siamese detail recovery network (FSD-Net) is designed in this paper. Firstly, from the perspective of frequency domain analysis, a theory on the dual roles of frequency domain components is introduced to reveal the robustness of low-frequency components to pseudo-changes and the dual semantic noise attributes of high-frequency components. Based on this theory, a frequency-aware context-guided difference (FCGD) module is designed. By explicitly decoupling the difference features into low-frequency global components and high-frequency residual components, it utilizes the prior low-frequency scene as a semantic gate to adaptively modulate the high-frequency differences, which effectively suppress pseudo-change interference. Subsequently, a detail recovery block (DRB), based on sub-pixel convolution, is constructed. This achieves unbiased spatial rearrangement through the semantic redundancy of channel dimensions, which avoids the checkerboard artifacts of traditional upsampling, and by employing a progressive multi-stage upsampling strategy to integrate shallow detail features from the encoder. The experimental results on the three public datasets of LEVIR-CD, WHU-CD, and CDD-CD demonstrate that the FSD-Net outperforms current mainstream methods (e.g., ChangeFormer, BAN, and so on) in core metrics such as F1 score and IoU, with a particularly significant improvement in recall. The ablation experiments validate the effectiveness and complementarity of the FCGD and DRB. Parameter sensitivity analysis indicates that the auxiliary loss weight λ is dataset dependent, with λ = 0.1 serving as a robust default choice. This study provides an efficient and reliable solution for change detection in high-resolution remote sensing imagery.