To address the issue of “domain shift”, researchers have proposed numerous single-domain generalization methods for crowd counting. These methods primarily focus on feature alignment to directly extract domain-invariant features, thereby enhancing the model's generalization ability. However, we argue that in single-source domain scenarios, where the style differences in video images are minimal, the domain-invariant features extracted directly from the entire image often contain a significant amount of redundant background information. Based on this observation, we propose a crowd counting network via single domain generalization (Crowd-SDG). Specifically, to increase the diversity of the source domain, we design a block-style attention strategy. This strategy focuses on different regions of the same image and combines data augmentation techniques such as lighting enhancement, cropping, and stretching to obtain more diverse features. To extract domain-invariant features that contain only crowd-related information, we introduce a feature recognition module. This module first uses a contrastive language image pre-training to obtain image patches that contain only crowd features, then extracts the domain-invariant features related to the crowd via similarity computation. Additionally, we design a dynamic feature reconstruction module to effectively integrate domain-invariant features with regular features. This not only enhances the representation of domain-invariant information but also reduces the loss of fine-grained details. Extensive experimental results demonstrate that our method exhibits superior generalization performance across multiple datasets.

With the continuous growth of the population, crowd counting plays a crucial role in intelligent monitoring systems for the Internet of Things (IoT) and smart city development. Accurate monitoring of crowd density not only helps maintain public safety but also effectively promotes the development of smart cities. Currently, supervised crowd counting techniques have made significant progress in improving accuracy, but these methods rely on expensive manual annotations and have limited generalization performance. To address these challenges, this article proposes an unsupervised crowd counting network based on contrastive learning, named CrowdCL. CrowdCL primarily leverages image-image contrastive learning and text-image contrastive learning to achieve unsupervised crowd counting. Specifically, in image-image contrastive learning, we strengthen the network’s ability to distinguish crowd features by designing progressive occlusion strategies and patch matching strategies, effectively differentiating crowd information from background information. In text-image contrastive learning, we construct ordered textual prompts to match ordered feature maps and use modality matching loss$(L_{m})$to guide the image encoder. Additionally, to reduce the loss of fine details and alleviate the interference of complex backgrounds, we design a coarse-grained filtering strategy during the testing phase, assigning higher weights to crowd patches with greater potential. Experiments on multiple public datasets show that CrowdCL not only achieves outstanding performance but also outperforms some fully supervised methods in cross-dataset testing.

Domain generalization (DG) has shown significant potential for cross-scene hyperspectral image (HSI) classification, wherein a model is trained exclusively on the source domain (SD) and can be directly transferred to an unseen target domain (TD). Current DG-based methods focus only on expanding the distribution of source domains by randomizing the style of the entire HSI cube. They fail to account for the domain shift problem caused by the variance of spatial land-cover distribution (context semantics), which results in SD-specific patterns being overly emphasized during training and, consequently, limiting the generalizability. Moreover, such randomization on cubes may introduce undesirable artifacts, such as blurring or distortion, leading to semantically compromised samples. In this paper, aFourier-based spectral-spatial generator(FSSG) is proposed to generate diversified and robust generative domain (GD). Specifically, a Fourier disentanglement is developed to construct spectral expansion (SpeE) and spatial expansion (SpaE) from pixel-wise and region-wise levels, respectively. In SpeE, the style information is transmitted across pixels in a privacy-protecting way, i.e., SD shares the semantic information with the GD. In SpaE, an effective continuous frequency space interpolation is employed to transmit the styles and semantics information across cubes, which enables GD to bridge inter-domain gaps in both styles and context semantics. To further alleviate the over-emphasis on SD-specific patterns, a relaxation procedure is integrated within an adversarial training based on a coarse-to-fine paradigm, which facilitates the HSI cubes to gain more robust context semantics. Extensive experiments and analyses, conducted with two baseline methods across three public datasets, demonstrate the superiority of the proposed approach.

Cross-scene hyperspectral image classification faces significant challenges due to spectral-spatial domain shifts. Existing domain generalization (DG) methods typically generate extended domains (EDs) to enhance the diversity of source domains (SDs) and extract domain-invariant features for target domain classification. However, these methods often struggle to balance the invariance and diversity of ED features. For instance, placing excessive emphasis on diversity can compromise cross-domain invariance. Additionally, they tend to overlook domain-specific features that are crucial for classification, which limits generalization performance. To address these challenges, this paper proposes an ensemble network based on Fourier phase preservation and spectral random augmentation (FSENet), which generates high-quality EDs and effectively integrates domain-invariant and domain-specific features to improve classification accuracy. FSENet consists of three key components: 1) the Fourier Spatial Augmentation Module, which preserves phase while perturbing amplitude to enhance both the invariance and diversity of spatial features; 2) the Random-augmented Spectral Channel Attention Module, which employs channel attention and random perturbations to boost the discriminative ability of spectral features; and 3) the Domain Label-guided Ensemble Framework, which utilizes a multi-branch architecture to combine domain-invariant and domain-specific features from multiple SDs, overcoming the limitations of single-feature representations. Experiments on four real-world hyperspectral datasets, including Houston, Pavia, HyRANK and WHU, demonstrate that FSENet achieves a cross-scene classification accuracy improvement of 2.27% – 4.74% compared to state-of-the-art domain adaptation and DG methods. Ablation studies confirm the effectiveness of each proposed module. The code of this work will be available at https://github.com/goesfor/FSENet.

Cross-domain few-shot hyperspectral image classification involves acquiring knowledge from a substantial set of labeled samples in the source domain and subsequently applying this knowledge to tasks within target domains characterized by a limited number of labeled samples. Few-shot learning combined with domain adaptation is currently a popular method for the above problem, but how to learn effective representations in the few-shot learning stage and the domain adaptation stage respectively is still a huge challenge. Therefore, we propose a new hyperspectral image classification method that adds the gradually vanishing bridge(GVB) module in the domain adaptation stage to improve domain-invariant representation learning, and adds the bias diminishing(BD) module in the few-shot learning stage to improve metric-based representation learning. The experimental results demonstrate that the proposed method outperforms the cited state-of-the-art methods on four public HSI datasets.

To handle noisy labels in hyperspectral image classification, this paper proposes a channel-grouped network combining with noise detection. It does not require heavy intervention on noisy data but directly handles the noise during the training process to enable the model to learn more accurate decision boundaries. Specifically, a Grouped Multi-scale Network (GMNet) is built to obtain scattered spatial–spectral representations and two attention mechanisms adapted to noise are proposed before and after the model body. Finally, leveraging the idea of noise detection, we determine sample credibility based on information from the nearest neighbors and propose a noise-robust loss function. We simultaneously utilize the feature information of the projection space and the coordinate information to get the nearest neighbors. GMNet+ as a combination of the proposed method and an additional contrastive learning framework is also investigated. Extensive experiments on two public datasets with various noise ratios demonstrate the effectiveness of our method.

Although existing Convolutional Neural Networks (CNNs) have exhibited commendable performance in hyperspectral image classification, they often emphasize local features. In recent years, transformers have garnered interest for capturing global features in hyperspectral images. This paper introduces a hyperspectral image classification framework, Dilated Convolution in CNN and Transformer Networks (DCCTnet), which employs various types of convolutions for local feature extraction and integrates dilated convolutions into multi-head self-attention mechanisms. Initially, multi-stage dilated convolutions are applied for multi-level feature extraction from the image. Following this, a grouped convolution is employed for feature fusion. To integrate CNN and transformers seamlessly, a pioneering Multi-Scale Multi-Head Self-Attention mechanism (MS-MHSA) is proposed, incorporating multiple dilated convolutions in MHSA to capture local-global multi-scale hyperspectral features. This mechanism is seamlessly integrated with the CNN branch, harnessing the strengths of both CNN and transformers. Through extensive experiments on two standard datasets, our proposed method demonstrates higher classification accuracy compared to other state-of-the-art networks.

Currently, existing semi-supervised crowd counting methods usually learn unlabeled images through pseudo-labels and spatial consistency regularization paradigm. However, due to extremely limited labeled data and noise in density map pseudo-labels, the counting performance of the model is greatly limited. Although multi-task learning can help the model improve its feature representation ability, it seriously ignores the importance of multi-task collaboration. Therefore, to overcome the above problems, we propose a multi-task pseudo-label self-correction (MTPS) framework for crowd counting, which combines different tasks to enhance the correlation between tasks and reduce training bias. For labeled data, multi-task collaboration enables the model to fully explore the potential information in limited samples; for unlabeled data, multi-task collaboration can obtain more accurate pseudo-labels than density estimation task. In addition, to effectively suppress the problem of inaccurate pseudo-labels caused by noise, we propose a pseudo-label self-correction strategy based on multi-task collaboration. This strategy starts from the perspective of task to task to gradually reduce the interference of background noise and obtain higher quality pseudo-labels. A large number of experiments on three public datasets show that the proposed MTPS achieves superior counting performance.

Hyperspectral image (HSI) classification is an important topic in remote sensing tasks, aiming to exploit the spectral and spatial information in HSI to assign category labels to each pixel. Recently, numerous deep learning (DL) methods such as convolutional neural network (CNN) and transformer, have been introduced and achieved promising results. These DL-based models tend to use patch-based input patterns, where a patch is extracted around a central pixel and the neighboring pixels serve as auxiliary roles. However, compared to developing novel models, characteristics of the patch-based HSI classification framework have not been adequately explored. Most existing patch-based DL methods tend to overlook the potential relationships between the central pixel and its neighboring pixels. In addition, the unique spectral characteristics of HSI (e.g. adjacent bands possess higher correlation) require special attention. Moreover, the design of feature extractors for HSIs appears to be constrained primarily to the spatial and spectral domains. To cope with the above problems, we propose an HSI classification specifically designed model (HSIC-SDM), which encompasses three branches—frequency feature learning, spatial feature learning and spectral feature learning. The frequency feature learning branch employs a Fourier-based feature extractor to acquire representations in the frequency domain. The spatial feature learning branch emphasizes the significance of central pixel and edge information. The spectral feature learning branch concerns the high correlations in the adjacent spectral bands. Multidomain features derived from three branches provide diverse perspectives for the final classification. Extensive experiments on three datasets demonstrate the superiority of the proposed method over the state-of-the-art (SOTA) compared methods. The code of our work will be available publicly at https://github.com/Tikiten/HSIC-SDM.

Presently, to obtain a more accurate density map and crowd number, existing methods often count by combining training RGB images and depth images. However, these methods are not ideal for capturing and fusing complementary features in RGB-D. Therefore, to solve the above problems, we propose a collaborative cross-modal attention network named CCANet for accurate RGB-D crowd counting. CCANet is mainly composed of the collaborative cross-modal attention module (CCAM) and the collaborative cross-modal fusion module (CCFM). Specifically, CCAM focuses on adaptive, interleaved RGB-D information through channel and spatial cross-modal attentions to fully capture complementary features in different modes. CCFM can adaptively integrate these features by weighing the importance of the above complementary features. A large number of experiments on the ShanghaiTechRGBD and MICC benchmarks have proven the effectiveness of CCANet in RGB-D crowd counting. In addition, our CCANet is generally applicable to multimodal crowd counting and has achieved superior counting performance on the RGBT-CC benchmark.

Crowd counting is a task of intelligent applications, and its operation efficiency is very important. However, in order to obtain a better counting performance, most of the existing works often design larger and more complex network structures, which will result in them occupying considerable memory, time and other resources at runtime, seriously limiting their deployment scope and making it difficult to be widely used in practical scenarios. In this paper, in order to overcome the above problems, we propose an effective lightweight encoding-decoding crowd counting network, named Lw-Count. Specifically, in the encoding process, we design an efficient and lightweight convolution module (ELCM), which extracts the crowd features in the network through a refined ghost block to reduce the network parameters and computing cost, and solves the problem of inaccurate counting caused by uneven crowd distribution through spatial group normalization (SGN). In the decoding process, we design a scale regression module (SRM) to reduce the error details and chessboard effect caused by linear interpolation and deconvolution. In addition, we design a new loss function, which enhances the spatial correlation of the density map and the sensitivity of the crowd through a regional normalized cross-correlation loss and counting loss, to ensure the counting accuracy. Extensive experiments on five mainstream datasets demonstrate that Lw-Count achieves an optimal trade-off between counting performance and running speed compared with other methods.

Convolutional neural networks (CNNs), a kind of feedforward neural network with a deep structure, are one of the representative methods in hyperspectral image (HSI) classification. However, redundant information and interclass interference are common and challenging problems in HSI classification. In addition, if the spectral and spatial information is not properly extracted and analyzed, it will affect the classification performance of the network to a great extent. Aiming at these issues, this article proposes an HSI classification method based on an adaptive hash attention mechanism and a lower triangular network (AHA-LT). First, the attention mechanism is introduced in the preprocessing stage, which is composed of the spectral attention module and the adaptive hash spatial attention module in series. Then, the data processed by the attention mechanism are introduced into the lower triangular network (LTNet) to obtain the fused high-dimensional semantic features. Finally, we compress the features and obtain the output classification results through several fully connected layers. Among them, LTNet is composed of 2-D–3-D CNN and multiscale features. The network integrates the characteristics of multibranch, feature fusion, feature compression, and skip connections. Extensive experiments on four widely used HSI data sets show that the proposed method can obtain a great improvement in performance compared with the existing methods.

Graph-based semisupervised hyperspectral image (HSI) classification methods have obtained extensive attention. In graph-based methods, a graph is first constructed, and then the label propagation is carried out on the constructed graph to obtain the labels for unknown samples. However, the results of label propagation may be unreliable, especially in the case of very limited labeled samples. To address the above problem, we propose dual sparse representation graph-based collaborative propagation (DSRG-CP) for HSI classification. Specifically, DSRG-CP adopts sparse representation (SR) to construct spectral and spatial graphs on spectral and spatial dimensions, respectively. Then, label propagation is performed on two graphs iteratively. In each iteration, only the samples with high classification confidence from one graph are added into another graph as labeled samples for the next label propagation. After several iterations, the labels of unlabeled samples are predicted by fusing the results of label propagation from two graphs. In addition, to make the spectral graph more discriminative, the regularizer of spectral statistical information is added into spectral SR model. To make the classification results more consistent in space, the superpixel block constraint is added into spatial graph model as regularizer. To evaluate the performance of the proposed method, DSRG-CP is compared with several graph-based methods and other state-of-the-art methods. Extensive experiments on real HSI data sets show that DSRG-CP can obtain competitive results for HSI classification.

In this paper, a new change detection method based on fully-connected conditional random field (FCCRF) with region potential is proposed. To deal with over-smoothing problem in FCCRF model, we propose to add region boundary constraint into FCCRF model. The proposed method defines the unary potential using the memberships of unsupervised fuzzy C-means clustering, designs the pairwise potential by a linear combination of Gaussian kernels using the complete set of pixels in the multi-temporal images to suppress noise effects, implements the region potential by the mean probability of pixels within image objects to preserve details of object boundary information. Experimental results demonstrate that the proposed method improves the change detection accuracy, turns out to be more robust against noise than traditional approaches.

An automatic change detection method based on conditional random field (CRF) is presented for high resolution remote sensing images in this paper. Marginalized denoising autoencoder is used to generate the difference image. The clustering results of Fuzzy C-means are applied to initialize the unary potentials of CRF. A scaled squared Euclidean distance between neighboring pixels in the observed images is introduced to define the pairwise potentials of CRF, which avoid training parameters and help improve the accuracy and the degree of automation. The experimental results obtained on different remote sensing images demonstrated the accuracy and efficiency of our proposed method.

This paper presents a new hyperspectral image (HSI) classification method which is capable of automatic feature learning while achieving high classification accuracy. The method contains two major modules: the spectral classification module and the spatial constraint module. Spectral classification module uses a deep network named stacked denoising autoencoders (SdA) to learn feature representation of the data. Through SdA, the data are projected nonlinearly from its original hyperspectral space to some higher dimensional space where more compact distribution is obtained. An interesting aspect of this method is that it does not need a feature design/extraction process guided by human prior. The suitable feature for the classification is learned by the deep network itself. Superpixel is utilized to generate the spatial constraints to refine the spectral classification results. By exploiting the spatial consistency of neighborhood pixels, the accuracy of classification is further improved by a big margin. Experiments on the public datasets reveal the superior performance of the proposed method.

In this paper, we propose an improved road detection method from high-resolution remote sensing images by means of a modified path opening and closing algorithm. To detect high curvature roads, we improved previously used adjacency graphs by enlarging the path angle of modified adjacency graphs. Furthermore, to remove noise efficiently, we combined the length of paths with road geometrical features to set up road extraction decision rules. A chief advantage of our approach is to be able to extract roads consisting of an unknown number of segments without manual parameter adjustment. We evaluated our proposed method systematically with a variety of images from IKONOS, EROS and aerial images. Experimental results demonstrated the accuracy and efficiency of our proposed method.

Fuzzy clustering algorithms have been widely used in brain magnetic resonance (MR) image segmentation. However, due to the existence of noise and intensity inhomogeneity, many segmentation algorithms suffer from limited accuracy. In this paper, we propose a fuzzy clustering algorithm with robust spatially constraint for accurate and robust brain MR image segmentation. A novel spatial factor is proposed by incorporating the spatial information amongst neighborhood pixels with a simple metric. A new weight factor, which utilizes the intensity information of the original image, is constructed to filter the posterior and prior probabilities in the spatial neighborhood. The proposed method can preserve more details and overcome the over-smoothing disadvantage. Finally, the fuzzy objective function is integrated with the bias field estimation model to overcome the intensity inhomogeneity in the image and segment the brain MR images. Experimental results demonstrate that the proposed algorithm can substantially improve the accuracy of brain MR image segmentation.

A novel two-stage level set evolution method for detecting man-made objects in aerial images is described. The method is based on a modified Mumford-Shah model and it uses a two-stage curve evolution strategy to get a preferable detection. It applies fractal error metric, developed by Cooper, et al. (1994) at the first curve evolution stage and adds additional constraint texture edge descriptor that is defined by using DCT (discrete cosine transform) coefficients on the image at the next stage. Man-made objects and natural areas are optimally differentiated by evolving the partial differential equation. The method artfully avoids selecting a threshold to separate the fractal error image, while an improper threshold often results in great segmentation errors. Experiments of the segmentation show that the proposed method is efficient.