In the field of object detection, deep learning models have achieved great success in recent years. Despite these advances, detecting small objects remains difficult. Most objects in aerial images have features that are a challenge for traditional object detection techniques, including small size, high density, high variability, and varying orientation. Previous approaches have used slicing methods on high-resolution images or feature maps to improve performance. However, existing slicing methods inevitably lead to redundant computation. Therefore, in this article we present a novel adaptive slicing method named ASAHI (Adaptive Slicing Aided Hyper Inference), which can dramatically reduce redundant computation using an adaptive slicing size. Specifically, ASAHI focuses on the number of slices rather than the slicing size, that is, it adaptively adjusts the slicing size to control the number of slices according to the image resolution. Additionally, we replace the standard non-maximum suppression technique with Cluster-DIoU-NMS due to its improved accuracy and inference speed in the post-processing stage. In extensive experiments, ASAHI achieves competitive performance on the VisDrone and xView datasets. The results show that the mAP50 is increased by 0.9% and the computation time is reduced by 20–25% compared with state-of-the-art slicing methods on the TPH-YOLOV5 pretrained model. On the VisDrone2019-DET-val dataset, our mAP50 result is 56.4% higher, demonstrating the superiority of our approach.

Recently, with the remarkable advancements of deep learning in the field of image processing, convolutional neural networks (CNNs) have garnered widespread attention from researchers in the domain of hyperspectral image (HSI) classification. Moreover, due to the high performance demonstrated by the transformer architecture in classification tasks, there has been a proliferation of neural networks combining CNNs and transformers for HSI classification. However, the majority of the current methods focus on extracting spatial–spectral features from the HSI data of a single size for a pixel, overlooking the rich multi-scale feature information inherent to the data. To address this problem, we designed a novel transformer network with a CNN-enhanced cross-attention (TNCCA) mechanism for HSI classification. It is a dual-branch network that utilizes different scales of HSI input data to extract shallow spatial–spectral features using a multi-scale 3D and 2D hybrid convolutional neural network. After converting the feature maps into tokens, a series of 2D convolutions and dilated convolutions are employed to generate two sets of Q (queries), K (keys), and V (values) at different scales in a cross-attention module. This transformer with CNN-enhanced cross-attention explores multi-scale CNN-enhanced features and fuses them from both branches. Experimental evaluations conducted on three widely used hyperspectral image (HSI) datasets, under the constraint of limited sample size, demonstrate excellent classification performance of the proposed network.

Deep learning has been widely used in remote sensing image segmentation, while a lack of training data remains a significant issue. The few-shot segmentation of remote sensing images refers to the segmenting of novel classes with a few annotated samples. Although the few-shot segmentation of remote sensing images method based on meta-learning can get rid of the dependence on large data training, the generalization ability of the model is still low. This work presents a few-shot segmentation of remote sensing images with a self-supervised background learner to boost the generalization capacity for unseen categories to handle this challenge. The methodology in this paper is divided into two main modules: a meta learner and a background learner. The background learner supervises the feature extractor to learning latent categories in the image background. The meta learner expands on the classic metric learning framework by optimizing feature representation through contrastive learning between target classes and latent classes acquired from the background learner. Experiments on the Vaihingen dataset and the Zurich Summer dataset show that our model has satisfactory in-domain and cross-domain transferring abilities. In addition, broad experimental evaluations on PASCAL-5ⁱ and COCO-20ⁱ demonstrate that our model outperforms the prior works of few-shot segmentation. Our approach surpassed previous methods by 1.1% with ResNet-101 in a 1-way 5-shot setting.

Semantic segmentation of high-resolution remote sensing images plays an important role in many practical applications, including precision agriculture and natural disaster assessment. With the emergence of a large number of studies on convolutional neural networks, the performance of the semantic segmentation model of remote sensing images has been dramatically promoted. However, many deep convolutional network models do not fully refine the segmentation result maps, and, in addition, the contextual dependencies of the semantic feature map have not been adequately exploited. This article proposes a hierarchical refinement residual network (HRRNet) to address these issues. The HRRNet mainly consists of ResNet50 as the backbone, attention blocks, and decoders. The attention block consists of a channel attention module (CAM) and a pooling residual attention module (PRAM) and residual structures. Specifically, the feature map output by the four blocks of Resnet50 is passed through the attention block to fully explore the contextual dependencies of the position and channel of the semantic feature map, and, then, the feature maps of each branch are fused step by step to realize the refinement of the feature maps, thereby improving the segmentation performance of the proposed HRRNet. Experiments show that the proposed HRRNet improves segmentation result maps compared with various state-of-the-art networks on Vaihingen and Potsdam datasets.

In this research, we present the Spatial-Aware Transformer (SAT), an enhanced implementation of the Swin Transformer module, purposed to augment the global modeling capabilities of existing transformer segmentation mechanisms within remote sensing. The current landscape of transformer segmentation techniques is encumbered by an inability to effectively model global dependencies, a deficiency that is especially pronounced in the context of occluded objects. Our innovative solution embeds spatial information into the Swin Transformer block, facilitating the creation of pixel-level correlations, and thereby significantly elevating the feature representation potency for occluded subjects. We have incorporated a boundary-aware module into our decoder to mitigate the commonly encountered shortcoming of inaccurate boundary segmentation. This component serves as an innovative refinement instrument, fortifying the precision of boundary demarcation. After these strategic enhancements, the Spatial-Aware Transformer achieved state-of-the-art performance benchmarks on the Potsdam, Vaihingen, and Aerial datasets, demonstrating its superior capabilities in recognizing occluded objects and distinguishing unique features, even under challenging conditions. This investigation constitutes a significant advancement toward optimizing transformer segmentation algorithms in remote sensing, opening a wealth of opportunities for future research and development.

Spectral unmixing is among one of the major hyperspectral image analysis tasks that aims to extract basic features (endmembers) at the subpixel level and estimate their corresponding proportions (fractional abundances). Recently, the rapid development of deep learning networks has provided us with a new method to solve the problem of spectral unmixing. In this paper, we propose a spatial-information-assisted spectral information learning unmixing network (SISLU-Net) for hyperspectral images. The SISLU-Net consists of two branches. The upper branch focuses on the extraction of spectral information. The input of the upper branch is a number of pixels randomly extracted from the hyperspectral image. The data are fed into the network as a random combination of different pixel blocks each time. The random combination of batches can boost the network to learn global spectral information. Another branch focuses on learning spatial information from the entire hyperspectral image and transmitting it to the upper branch through the shared weight strategy. This allows the network to take into account the spectral information and spatial information of HSI at the same time. In addition, according to the distribution characteristics of endmembers, we employ Wing loss to solve the problem of uneven distributions of endmembers. Experimental results on one synthetic and three real hyperspectral data sets show that SISLU-Net is effective and competitive compared with several state-of-the-art unmixing algorithms in terms of the spectral angle distance (SAD) of the endmembers and the root mean square error (RMSE) of the abundances.