Recently, there are many works on discriminant analysis, which promote the robustness of models against outliers by using L₁- or L2,1-norm as the distance metric. However, both of their robustness and discriminant power are limited. In this article, we present a new robust discriminant subspace (RDS) learning method for feature extraction, with an objective function formulated in a different form. To guarantee the subspace to be robust and discriminative, we measure the within-class distances based on L2,s-norm and use L2,p-norm to measure the between-class distances. This also makes our method include rotational invariance. Since the proposed model involves both L2,p-norm maximization and L2,s-norm minimization, it is very challenging to solve. To address this problem, we present an efficient nongreedy iterative algorithm. Besides, motivated by trace ratio criterion, a mechanism of automatically balancing the contributions of different terms in our objective is found. RDS is very flexible, as it can be extended to other existing feature extraction techniques. An in-depth theoretical analysis of the algorithm's convergence is presented in this article. Experiments are conducted on several typical databases for image classification, and the promising results indicate the effectiveness of RDS.

Most existing RGB-D salient object detection (SOD) methods are primarily focusing on cross-modal and cross-level saliency fusion, which has been proved to be efficient and effective. However, these methods still have a critical limitation, i.e., their fusion patterns – typically the combination of selective characteristics and its variations, are too highly dependent on the network’s non-linear adaptability. In such methods, the balances between RGB and D (Depth) are formulated individually considering the intermediate feature slices, but the relation at the modality level may not be learned properly. The optimal RGB-D combinations differ depending on the RGB-D scenarios, and the exact complementary status is frequently determined by multiple modality-level factors, such as D quality, the complexity of the RGB scene, and degree of harmony between them. Therefore, given the existing approaches, it may be difficult for them to achieve further performance breakthroughs, as their methodologies belong to some methods that are somewhat less modality sensitive. To conquer this problem, this paper presents the Modality-aware Decoder (MaD). The critical technical innovations include a series of feature embedding, modality reasoning, and feature back-projecting and collecting strategies, all of which upgrade the widely-used multi-scale and multi-level decoding process to be modality-aware. Our MaD achieves competitive performance over other state-of-the-art (SOTA) models without using any fancy tricks in the decoder’s design. Codes and results will be publicly available at https://github.com/MengkeSong/MaD .

In this paper, a spectral-spatial convolution neural network with Siamese architecture (SSCNN-S) for hyperspectral image (HSI) change detection (CD) is proposed. First, tensors are extracted in two HSIs recorded at different time points separately and tensor pairs are constructed. The tensor pairs are then incorporated into the spectral-spatial network to obtain two spectral-spatial vectors. Thereafter, the Euclidean distances of the two spectral-spatial vectors are calculated to represent the similarity of the tensor pairs. We use a Siamese network based on contrastive loss to train and optimize the network so that the Euclidean distance output by the network describes the similarity of tensor pairs as accurately as possible. Finally, the values obtained by inputting all tensor pairs into the trained model are used to judge whether a pixel belongs to the change area. SSCNN-S aims to transform the problem of HSI CD into a problem of similarity measurement for tensor pairs by introducing the Siamese network. The network used to extract tensor features in SSCNN-S combines spectral and spatial information to reduce the impact of noise on CD. Additionally, a useful four-test scoring method is proposed to improve the experimental efficiency instead of taking the mean value from multiple measurements. Experiments on real data sets have demonstrated the validity of the SSCNN-S method.

Change detection (CD) is a hot issue in the research of remote sensing technology. Hyperspectral images (HSIs) greatly promote the development of CD technology because of their high resolution in the spectral domain. However, some traditional CD methods currently applied to low-dimensional and multispectral images cannot adapt to the complex high-dimensional features of the HSIs. In addition, the spectral measurements of the HSI contain a lot of noise and redundancy, which greatly contaminates spectral-only information for CD. In order to fully extract the discriminant features of HSI to improve the accuracy of CD, this article proposes a three-directions spectral-spatial convolution neural network (TDSSC). A novel method for three-direction decomposition of hyperspectral change tensors is proposed-change tensor is decomposed along the spectral direction and two spatial directions to get a single tensor containing the spectral information and two kinds of tensors containing the spectral-spatial information. TDSSC uses 1-D convolution to extract spectral features from the spectral direction as well as reducing the tensor dimension, which helps the latter network to be lightweight and significantly improves the speed of change detection. Also, it uses 2-D convolution to extract spectral-spatial features from two spatial directions of the reduced tensor, and to extract features from different directions to improve the accuracy and Kappa value of CD. The experimental results of three real hyperspectral datasets show that TDSSC is superior to most existing CD methods.

This paper advocates a novel learning solution to the modeling of long-term spatial-temporal saliency consistency in order to boost the accuracy for video saliency detection. Conventional methods typically utilize the “slack” spatial-temporal model to locally ensure the smoothness of the computed video saliency, yet they could easily encounter the performance tradeoff dilemma (i.e., detection’ accuracy and integrity). In contrast, our novel approach proposes the bi-level learning strategy to globally exploit the saliency consistency while overcoming the above difficulty. Our method first starts with the contrast computation of low-level saliency clues in a frame-wise manner. Then, based on such obtained saliency clues, we devise a novel bilevel Markov Random Field (bMRF) solution to conduct semantic labelling, which can explicitly indicates both the salient salient foregrounds and non-salient nearby surroundings with high confidence while shrinking the low confidence remains. In such a way, the spatial-temporal consistency constraint is embedded intrinsically into the above explicit semantic labels, and we avoid the performance trade-off problem from occurring. Next, based on those semantic labels made by our bMRF method, we further propose to learn multiple non-linear feature transformations to enlarge the feature margin between the salient foregrounds and the non-salient nearby surroundings, whose key rationale is to resort to long-term common consistencies to enforce the spatial-temporal smoothness. Thus, we can utilize these learned non-linear feature transformations to simultaneously suppress those short-term false-alarms and correct those hollow effects. To validate our new approach, we conduct extensive experiments on 5 publicly available benchmarks, and make comprehensive, quantitative evaluations between our method and 17 state-ofthe- art techniques. All the results demonstrate our method’s advantages in terms of accuracy, reliability, robustness, and versatility.