ABSTRACT In urban road scenes, the cross‐domain data distribution differences caused by light and weather changes make the generalization performance of single‐domain trained object detectors in unknown weather scenarios significantly degraded (e.g., daytime‐sunny trained models in dusk‐rainy scenarios reduce the detection accuracy by more than 46% on average); moreover, faster R‐CNN suffers from insufficient generalization capability and inefficient real‐time inference due to architectural constraints. To address the above challenges, this paper proposes the CEN‐RTDETR method to improve the single‐domain generalization capability through a collaborative enhancement strategy: CP‐Mix color channel permutation dynamically simulates the RGB color channel permutation to simulate the multi‐sky color bias phenomenon, and enhances the color robustness of the input data; NP feature normalization perturbation applies a random feature perturbation to the channel statistics of the shallow feature map to optimize the extraction of texture, color and other basic style features; CORAL Loss minimizes the difference in feature distribution between the source domain and the virtual target domain through second‐order statistical matching. The experimental results show that CEN‐RTDETR achieves significant performance improvement on the cross‐weather scenario dataset DWD: The mean average precision ([email protected]) across different weather scenarios increases from 39.66% to 42.72% (+3.06%), especially for dusk‐rainy and night‐rainy scenarios, where [email protected] rises from 32.9% and 18.6% to 39.0% (+6.1%) and 24.1% (+5.5%) in extreme weather. The method in this paper effectively solves the cross‐domain generalization and efficiency problems in single‐domain generalized object detection, which provides new technical possibilities for real‐time detection in complex urban road scenes.

Timely detection and treatment of road cracks are crucial to prevent further deterioration of pavement. An accurate road crack detection algorithm can significantly reduce the human resources required for pavement maintenance. However, existing CNN-based object detectors, such as the You Only Look Once (YOLO) series of algorithms, face challenges such as receptive field fixation and information loss during feature extraction, resulting in lower accuracy in road crack detection. Therefore, we propose the Selective Dynamic Feature Compensation-YOLO (SDFC-YOLO) algorithm for pavement distress detection. Firstly, we introduce the Dynamic Downsampling Module (DDM), which adaptively adjusts the sampling positions of the convolutional kernel during the feature extraction process, addressing the issue of a fixed receptive field. Secondly, we propose a novel feature fusion method compensating for lost feature information in the path aggregation network. Lastly, we design a multi-scale weight selection module based on the above feature fusion method. It aims to utilize the channel weights of high-level features to guide bottom-level features and select more important features for compensation, thereby further enhancing detection accuracy. Experimental results demonstrate that compared to the benchmark model YOLOv8s, our method improves Precision (P), Recall (R), mean Average [email protected] ([email protected]), and F1 score on the UAPD dataset by 6.8%, 2.4%, 5.3%, and 4.3%, respectively. Similarly, on the UAV-PDD2023 dataset, the aforementioned metrics are enhanced by 3.1%, 4.4%, 2.7%, and 3.8%, respectively. Furthermore, our method takes only 10.53ms to process a$1280\times 1280$resolution image, which fully meets the requirement of real-time detection.

Band selection is a valid method to reduce redundant information in hyperspectral images (HSIs). Typically, two methods are used to select representative bands: ranking bands based on predefined criteria and selecting cluster centers by grouping bands. It is advantageous to combine these two methods for hyperspectral band selection tasks since their benefits are complimentary. To take full of these advantages, we propose a hyperspectral band selection method using curve-fitting subspace partition (CFSP), including the subspace partition method and local context representative band selection method. The contributions of this letter are summarized below: 1) through fitting the spectral curves and then utilizing the point of maximum curvature to partition the band set, the similar and adjacent bands can be divided into the same group, which is very consistent with the way of subspace partition and 2) a representative band selection method in the local context is proposed. The locally optimal bands are selected sequentially to constitute the candidate band set. Then, through ranking and iteratively updating the candidate band set, we can effectively find the desired bands. The experiments on three public HSI datasets show that the proposed method has significant advantages compared with some advanced competitors. In particular, on the Salinas dataset, the selected bands achieved an excellent average overall accuracy (OA) of 91.32% using the support vector machine (SVM) classifier.

Hyperspectral band selection is viewed as an effective dimension reduction method. Recently, researchers present graph-based clustering for hyperspectral image (HSI) processing. However, most of them conduct clustering on a fixed data matrix so that it is sensitive to the quality of initial matrix. Moreover, these algorithms apply spectral clustering to obtain the final clustering result in increasing the time consumption. Based on these facts, we propose a block diagonal representation learning algorithm (BDRLA) in this paper. BDRLA generates a high-quality similarity matrix by approximating the initial affinity matrix. Meanwhile, motivated to the spectral bands distance similarity matrix has a clear diagonal structure, a block diagonal similarity matrix with ordered partition points based upon the ℓ2-norm is constructed. By doing so, the obtained similarity matrix is directly applied to subsequent processing without extracting the clustering indicators. Additionally, in order to estimate the importance of bands, dictionary learning is adopted to select the representative band in each cluster. Extensive experiment results on three public datasets indicate that the bands selected by the proposed method achieve satisfactory performance.

Various methods are proposed to reduce the dimensions of hyperspectral image by band selection in recent years. Most methods select one band from each group to construct a band subset. However, the redundancy in the selected bands from different groups is neglected. Furthermore, the researchers do not pay enough attention to how many bands are appropriated for selection. To solve these issues, we propose a hyperspectral band selection method via difference between inter-groups (DIG), which includes grouping strategy and ranking strategy. Specifically, the grouping strategy adopts intra-group similarity to reasonably distribute all partitioning point positions. The similarity of bands within the same group is significantly improved. For the ranking strategy, it not only takes into account the knowledge and intra-group similarity of bands, but also evaluates the differences between each band and other inter-group bands. The redundancy in band subset is reduced sufficiently. In order to accurately obtain the optimal number of bands, an evaluation function is designed to measure the information content and redundancy in various band subsets. Experimental results from different aspects show that the proposed model has a large performance advantage on three public datasets.