Unmanned aerial vehicle (UAV) remote-sensing images present unique challenges to the object-detection task due to uneven object densities, low resolution, and drastic scale variations. Downsampling is an important component of deep networks that expands the receptive field, reduces computational overhead, and aggregates features. However, object detectors using multi-layer downsampling result in varying degrees of texture feature loss for various scales in remote-sensing images, degrading the performance of multi-scale object detection. To alleviate this problem, we propose a lightweight texture reconstructive downsampling module called TRD. TRD models part of the texture features lost as residual information during downsampling. After modeling, cascading downsampling and upsampling operators provide residual feedback to guide the reconstruction of the desired feature map for each downsampling stage. TRD structurally optimizes the feature-extraction capability of downsampling to provide sufficiently discriminative features for subsequent vision tasks. We replace the downsampling module of the existing backbone network with the TRD module and conduct a large number of experiments and ablation studies on a variety of remote-sensing image datasets. Specifically, the proposed TRD module improves 3.1% AP over the baseline on the NWPU VHR-10 dataset. On the VisDrone-DET dataset, the TRD improves 3.2% AP over the baseline with little additional cost, especially the ${\mathrm{AP}}_{\mathrm{S}}$, ${\mathrm{AP}}_{\mathrm{M}}$, and ${\mathrm{AP}}_{\mathrm{L}}$ by 3.1%, 8.8%, and 13.9%, respectively. The results show that TRD enriches the feature information after downsampling and effectively improves the multi-scale object-detection accuracy of UAV remote-sensing images.

In resource-constrained unmanned aerial vehicle (UAV) infrared image transmission, infrared small targets usually occupy only a small number of pixels, while conventional uniform resource allocation strategies fail to consider the semantic differences among image regions, resulting in inefficient resource utilization and loss of target-related information. To address this issue, this paper proposes a semantic-aware resource allocation method for infrared small-target detection. First, infrared images are divided into multiple grid cells, and the semantic importance of each cell is estimated based on preliminary detection results. Then, the cells are classified into different semantic levels, followed by differentiated bit allocation and region-wise reconstruction. Finally, a joint optimization problem of grid partitioning and resource allocation parameters is formulated and solved by the covariance matrix adaptation evolution strategy (CMA-ES) to optimize both detection performance and image reconstruction quality. Experimental results demonstrate that the proposed method achieves mIoU values of 0.6106 and 0.7445 at available bit rates of 12.5% and 50%, respectively, showing significant improvements over comparison methods. Moreover, CMA-ES obtains stable optimization results with fewer evaluations and provides a favorable balance between detection accuracy and reconstruction quality. These results indicate that the proposed method can effectively preserve target-related semantic information under limited communication resources, providing an effective solution for efficient UAV infrared image transmission and reliable small-target detection.

Accurately identifying navigable areas in wild autonomous driving environments using onboard visual sensors remains a challenging task due to highly complex and dynamic terrain. Existing methods often struggle to simultaneously balance global discrimination and local consistency, leading to insufficient terrain understanding. To address this, we propose the Navigable Terrain Segmentation Network (NTSNet), which integrates two core modules: the Gaussian Fourier Attention Module (GFAM) and the Entropy-Aware Fusion Module (EAFM). The GFAM introduces Gaussian Fourier-driven spatial priors to model long-range dependencies, enabling robust extraction of deep global semantic features and enhancing global discrimination. The EAFM, comprising the attention entropy modulation Unit and weight entropy aggregation Unit, leverages Shannon entropy to uncover latent advantageous information among neighboring features and adaptively fuse multi-scale representations, improving local consistency. Quantitative and qualitative experiments on two wild terrain datasets demonstrate that the proposed NTSNet achieves competitive segmentation performance against state-of-the-art methods, providing reliable pixel-level terrain perception for wild autonomous driving systems. The code for NTSNet will be made available at: https://github.com/lv881314/NTSNet .

Segmentation of unstructured natural environments is crucial for scene parsing in outdoor autopilot systems. However, directly applying existing segmentation networks often results in decreased accuracy, as these networks struggle to effectively handle external factors like uneven terrains and lighting variations. To tackle this issue, we propose a memory-based encoder-decoder network to improve the segmentation performance of existing networks. Specifically, we introduce a memory module that refines features from the encoder, comprising a read and a write memory module. The read memory module introduces random memory classes to identify various category semantic features from the encoder and refines these features using addressing weights. The write memory module, embedded within the read memory module, updates the random memory classes to better align with the encoder’s features. To optimize semantic features with memory classes, we employ a semi-hard triplet loss in the write memory module to reduce redundant memory classes. Our proposed method, tested on two challenging datasets for unstructured natural scene parsing, demonstrates that the introduced memory module significantly enhances the performance of existing segmentation networks while maintaining low computational costs.