Fine-grained object detection (FGOD) fundamentally comprises two primary tasks: object detection and fine-grained classification. In natural scenes, most FGOD methods benefit from higher instance resolution and fewer environmental variation, attributing more commonly associated with the latter task. In this paper, we propose a unified paradigm named Detector with Classifier2 (DC2), which provides a holistic paradigm by explicitly considering the end-to-end integration of object detection and fine-grained classification tasks, rather than prioritizing one aspect. Initially, our detection sub-network is restricted to only determining whether the proposal is a coarse-category and does not delve into the specific sub-categories. Moreover, in order to reduce redundant pixel-level calculation, we propose an instance-level feature enhancement (IFE) module to model the semantic similarities among proposals, which poses great potential for locating more instances in remote sensing images (RSIs). After obtaining the coarse detection predictions, we further construct a classification sub-network, which is built on top of the former branch to determine the specific sub-categories of the aforementioned predictions. Importantly, the detection network is performed on the complete image, while the classification network conducts secondary modeling for the detected regions. These operations can be denoted as the global contextual information and local intrinsic cues extractions for each instance. Therefore, we propose a multi-stream feature aggregation (MSFA) module to integrate global-stream semantic information and local-stream discriminative cues. Our whole DC2 network follows an end-to-end learning fashion, which effectively excavates the internal correlation between detection and fine-grained classification networks. We evaluate the performance of our DC2 network on two benchmarks SAT-MTB and HRSC2016 datasets. Importantly, our method achieves the new state-of-the-art results compared with recent works (approximately 7% mAP gains on SAT-MTB) and improves baseline by a significant margin (43.2% $v.s.~36.7$ %) without any complicated post-processing strategies. Source codes of the proposed methods are available at https://github.com/zhengshangdong/DC2.

Typical high-level vision tasks in hyperspectral image (HSI) processing, such as target detection, often suffer from insufficient information inherent in real-world sampled data. Super-resolution, a powerful tool in HSI low-level vision, is expected to enhance the accuracy of detection results by computationally providing the high-resolution HSI with additional information. However, existing solutions for HSI super-resolution and target detection have always been implemented independently. This conventionally adopted paradigm overlooks the interconnectedness between low-level and high-level visions, inevitably introducing additional errors, redundancies, and inefficiencies. To address this challenge, in this study, we put our efforts into exploring the uncharted continent of hyperspectral remote sensing, that is, realizing the mutual guidance and joint optimization of HSI super-resolution and target detection concurrently within a unified framework. Technically, we first construct different spectral bases to span the target and background subspaces of the underlying high-resolution HSI. Then, we look in-depth at the intrinsic properties of the HSI tensor, henceforth jointly optimizing both tasks by innovatively developing a novel low-cubic-rank tensor approximation model with a unique constrained energy minimization loss. While we have developed efficient algorithms to optimize the proposed model, we also put into place a refinement procedure for spectral bases, aimed at further enhancing the spectral fidelity of the fused results and the compact representation of the target subspace. Finally, empirical studies conducted on synthetic and real-world datasets substantiate that compared with state-of-the-art solutions, the proposed method delivers highly competitive and practical performance in terms of both tasks. Source codes are available at https://github.com/CX-He/HySRTD.git.

Hyperspectral image (HSI) target detection plays a pivotal role in both military and civilian sectors. Nevertheless, this task is fraught with challenges because of the limited availability of target samples and the intricate nature of the background within real-world HSIs. In this study, we present an innovative background learning model based on the orthogonal subspace-guided variational autoencoder, tailored to discern the background distribution in hyperspectral imagery. Given the scarcity of target samples, our model is exclusively trained on background spectral samples, enabling precise modeling of the background distribution. The crux of our approach lies in detecting disparities between the reconstructed HSI and the original HSI, providing a mechanism for faithful target identification. To procure background samples, a coarse detection of the test HSI is first conducted. However, this process proves challenging, as obtaining pristine background pixels is a formidable task. To mitigate the influence of suspicious target samples on the background reconstruction, we employ orthogonal subspace loss on the reconstructed HSI. Extensive experiments conducted on four real-world HSIs substantiate that the proposed framework performs highly competitively and the results outperform other state-of-the-art HSI target detection methods. The source codes of this study are available at https://github.com/CX-He/OS-VAE.

The recently proposed high-order tensor algebraic framework generalizes the tensor singular value decomposition (t-SVD) induced by the invertible linear transform from order-3 to order-d ( ). However, the derived order-d t-SVD rank essentially ignores the implicit global discrepancy in the quantity distribution of non-zero transformed high-order singular values across the higher modes of tensors. This oversight leads to suboptimal restoration in processing real-world multi-dimensional visual datasets. To address this challenge, in this study, we look in-depth at the intrinsic properties of practical visual data tensors, and put our efforts into faithfully measuring their high-order low-rank nature. Technically, we first present a novel order-d tensor rank definition. This rank function effectively captures the aforementioned discrepancy property observed in real visual data tensors and is thus called the discrepant t-SVD rank. Subsequently, we introduce a nonconvex regularizer to facilitate the construction of the corresponding discrepant t-SVD rank minimization regime. The results show that the investigated low-rank approximation has the closed-form solution and avoids dilemmas caused by the previous convex optimization approach. Based on this new regime, we meticulously develop two models for typical restoration tasks: high-order tensor completion and high-order tensor robust principal component analysis. Numerical examples on order-4 hyperspectral videos, order-4 color videos, and order-5 light field images substantiate that our methods outperform state-of-the-art tensor-represented competitors. Finally, taking a fundamental order-3 hyperspectral tensor restoration task as an example, we further demonstrate the effectiveness of our new rank minimization regime for more practical applications. The source codes of the proposed methods are available at https://github.com/CX-He/DTSVD.git.

Recently, tensor singular value decomposition (t-SVD) has demonstrated excellent performance in various high-dimensional information processing applications. However, in adapting t-SVD to handle the typical tensor data restoration tasks, such as hyperspectral image (HSI) denoising, the following questions remain inadequately addressed: 1) The existing tensor nuclear norm minimization (TNN) regime treats all tensor singular values alike; thus, it lacks flexibility and dominance in dealing with the sophisticated HSI tensor. 2) The existing t-SVD-based denoising methods can not directly process order-p(p> 3) tensors; thus, they fail to comprehensively exploit the high-dimensional structural correlation of the HSI tensor along different modes. To address the above challenges, in this study, we first generalize a novel weighted order-pTNN minimization regime, which integrates the adaptively reweighting strategy for matrix, third-order, and order-ptensors in a unified architecture. Subsequently, an efficient subspace low-rank learning model is established, using HSI denoising tasks as an application example to corroborate the superiority of the proposed regime in approximating the high-dimensional low-rank structure of natural tensor data. Extensive experimental results substantiate that our effort surpasses existing state-of-the-art low-rank tensor recovery methods in both restoration accuracy and efficiency. The source code is available at https://github.com/CX-He/WTNN.git.

Low-rank tensor representation philosophy has enjoyed a reputation in many hyperspectral image (HSI) low-level vision applications, but previous studies often failed to comprehensively exploit the low-rank nature of HSI along different modes in low-dimensional subspace, and unsurprisingly handled only one specific task. To address these challenges, in this paper, we figured out that in addition to the spatial correlation, the spectral dependency of HSI also implicitly exists in the coefficient tensor of its subspace, this crucial dependency that was not fully utilized by previous studies yet can be effectively exploited in a cascaded manner. This led us to propose a unified subspace low-rank learning regime with a new tensor cascaded rank minimization, named STCR, to fully couple the low-rankness of HSI in different domains for various low-level vision tasks. Technically, the high-dimensional HSI was first projected into a low-dimensional tensor subspace, then a novel tensor low-cascaded-rank decomposition was designed to collapse the constructed tensor into three core tensors in succession to more thoroughly exploit the correlations in spatial, nonlocal, and spectral modes of the coefficient tensor. Next, difference continuity-regularization was introduced to learn a basis that more closely approximates the HSI's endmembers. The proposed regime realizes a comprehensive delineation of the self-portrait of HSI tensor. Extensive evaluations conducted with dozens of state-of-the-art (SOTA) baselines on eight datasets verified that the proposed regime is highly effective and robust to typical HSI low-level vision tasks, including denoising, compressive sensing reconstruction, inpainting, and destriping. The source code of our method is released at https://github.com/CX-He/STCR.git.

With the rapid advancement of spectrometers, the imaging range of the electromagnetic spectrum starts growing narrower. The reduction of electromagnetic wave energy received in a single wavelength range leads more complex noise into the generated hyperspectral image (HSI), thus causing a severe cripple in the accuracy of subsequent applications. The requirement for the HSI mixed denoising algorithm’s accuracy is further lifted. To address this challenge, in this letter, we propose a novel difference continuity-regularized nonlocal tensor subspace low-rank learning (named DNTSLR) method for HSI mixed denoising. Technically, the original high-dimensional HSI data was first projected into a low-dimensional subspace spanned by a spectral difference continuous basis instead of an orthogonal basis, so the data continuity of the restored HSI spectrum and tensor low-rankness was guaranteed. Then, a cube matching strategy was employed to stack the nonlocal tensor patches from the projected coefficient tensor, and a shrinkage algorithm was used to approximate the low-rank coefficient tensor. Eventually, the subspace low-rank learning algorithm was designed to alternately separate the noise tensor and restore the latent clean low-rank HSI tensor. Extensive experiments on multiple open datasets validate that the proposed method realizes the state-of-the-art denoising accuracy for HSI.

In this letter, using the sparse unmixing framework, a weighted collaborative sparse and$L_{1/2}$low-rank regularization with superpixel segmentation method is proposed for hyperspectral unmixing. The method outlined here first uses superpixel segmentation to obtain local homogeneous regions. The reason for this approach is that the shape and size of superpixels are adaptive, which are better for obtaining homogeneous regions than square patches. Next, the weighted collaborative sparse term and$L_{1/2}$low-rank regularization were utilized to exploit the spatial and spectral correlation of each superpixel. In addition, the smoothness between adjacent pixels is enforced by total variation regularization. Finally, the proposed method and several state-of-the-art methods were tested on two simulated data sets and two real data sets. The results demonstrate the superiority of the method proposed here.

Currently, edge computing (EC), emerging as a burgeoning paradigm, is powerful in handling real-time resource provision for Internet of Things (IoT) applications. However, due to the spatial distribution of geographically sparse IoT devices and the resource limitations of EC units (ECUs), the resource utilization of corresponding edge servers is relatively insufficient and the execution performance is ineffective to some extent. A privacy leakage, including personal information, location, media data, etc., during the transmission process from IoT devices to edge servers severely restricts the application of ECUs in IoT. To address these challenges, a two-phase offloading optimization strategy is put forward for joint optimization of offloading utility and privacy in EC enabled IoT. Technically, a utility-aware task offloading method, named UTO, is devised first to obtain the goal of maximizing the resource utilization of ECUs and minimizing the implementation time cost. Then a joint optimization method, named JOM, for utility and privacy tradeoffs is designed to balance the privacy preservation and execution performance. Eventually, the experimental evaluations are designed to illustrate the efficiency and reliability of UTO and JOM.