Multispectral Scene Classification via Cross-Modal Knowledge Distillation

Scene change detection is the process of identifying the differences between the multi-temporal image scenes at the semantic level, which has significant potential in the application of urban development and land management. In this paper, we propose a novel deep convolution canonical correlation analysis neural network (DCCANet) architecture, which could consider the spectral-spatial-temporal correlation feature for scene change detection in remote sensing images. For this purpose, we put together the convolutional neural networks (CNNs) and the deep canonical correlation analysis (DCCA) into the end-to-end network. The CNN could get the spectral-spatial feature information for scene representation, while the later could enhance the temporal correlation by nonlinear high- dimensional transformations between the multi-temporal image scenes for scene change detection. Experiments with high-resolution remote sensing image scene datasets demonstrated that our proposed approach can get a better performance in scene classification and change detection.

Scene classification of high-resolution remote sensing images (HRRSI) is one of the most important means of land-cover classification. Deep learning techniques, especially the convolutional neural network (CNN) have been widely applied to the scene classification of HRRSI due to the advancement of graphic processing units (GPU). However, they tend to extract features from the whole images rather than discriminative regions. The visual attention mechanism can force the CNN to focus on discriminative regions, but it may suffer from the influence of intra-class diversity and repeated texture. Motivated by these problems, we propose an attention-based deep feature fusion (ADFF) framework that constitutes three parts, namely attention maps generated by Gradient-weighted Class Activation Mapping (Grad-CAM), a multiplicative fusion of deep features and the center-based cross-entropy loss function. First of all, we propose to make attention maps generated by Grad-CAM as an explicit input in order to force the network to concentrate on discriminative regions. Then, deep features derived from original images and attention maps are proposed to be fused by multiplicative fusion in order to consider both improved abilities to distinguish scenes of repeated texture and the salient regions. Finally, the center-based cross-entropy loss function that utilizes both the cross-entropy loss and center loss function is proposed to backpropagate fused features so as to reduce the effect of intra-class diversity on feature representations. The proposed ADFF architecture is tested on three benchmark datasets to show its performance in scene classification. The experiments confirm that the proposed method outperforms most competitive scene classification methods with an average overall accuracy of 94% under different training ratios.

Remote sensing scene classification is an important technology, which is widely used in military and civil applications. However, it is still a challenging problem due to the complexity of scene images. Recently, the development of remote sensing satellite and sensor devices has greatly improved the spatial resolution and semantic information of remote sensing images. Therefore, we propose a novel remote sensing scene classification approach to enhance the performance of scene classification. First, a spatial and channel-wise attention module is proposed to adequately utilize the spatial and feature information. Compare with other methods, channel-wise module works on the feature maps with diverse levels and pays more attention to semantic-level features. On the other hand, spatial attention module promotes correlation between foreground and classification result. Second, a novel classifier named broad-classifier is designed to enhance the discriminability. It greatly reduces the cost of computing in the meantime by broad learning system. The experimental results have show that our classification method can effectively improve the average accuracies on remote sensing scene classification data sets.

While convolutional neural networks (CNN) have been excellent for object recognition, the greater spatial variability in scene images typically meant that the standard full-image CNN features are suboptimal for scene classification. In this paper, we investigate a framework allowing greater spatial flexibility, in which the Fisher vector (FV) encoded distribution of local CNN features, obtained from a multitude of region proposals per image, is considered instead. The CNN features are computed from an augmented pixel-wise representation comprising multiple modalities of RGB, HHA and surface normals, as extracted from RGB-D data. More significantly, we make two postulates: (1) component sparsity - that only a small variety of region proposals and their corresponding FV GMM components contribute to scene discriminability, and (2) modal non-sparsity - within these discriminative components, all modalities have important contribution. In our framework, these are implemented through regularization terms applying group lasso to GMM components and exclusive group lasso across modalities. By learning and combining regressors for both proposal-based FV features and global CNN features, we were able to achieve state-of-the-art scene classification performance on the SUNRGBD Dataset and NYU Depth Dataset V2.

Acoustic scene classification is the task of identifying the scene from which the audio signal is recorded. Convolutional neural network (CNN) models are widely adopted with proven successes in acoustic scene classification. However, there is little insight on how an audio scene is perceived in CNN, as what have been demonstrated in image recognition research. In the present study, the Class Activation Mapping (CAM) is utilized to analyze how the log-magnitude Mel-scale filter-bank (log-Mel) features of different acoustic scenes are learned in a CNN classifier. It is noted that distinct high-energy time-frequency components of audio signals generally do not correspond to strong activation on CAM, while the background sound texture are well learned in CNN. In order to make the sound texture more salient, we propose to apply the Difference of Gaussian (DoG) and Sobel operator to process the log-Mel features and enhance edge information of the time-frequency image. Experimental results on the DCASE 2017 ASC challenge show that using edge enhanced log-Mel images as input feature of CNN significantly improves the performance of audio scene classification.

Semantic-level land-use scene classification is a challenging problem, in which deep learning methods, e.g., convolutional neural networks (CNNs), have shown remarkable capacity. However, a lack of sufficient labeled images has proved a hindrance to increasing the land-use scene classification accuracy of CNNs. Aiming at this problem, this paper proposes a CNN pre-training method under the guidance of a human visual attention mechanism. Specifically, a computational visual attention model is used to automatically extract salient regions in unlabeled images. Then, sparse filters are adopted to learn features from these salient regions, with the learnt parameters used to initialize the convolutional layers of the CNN. Finally, the CNN is further fine-tuned on labeled images. Experiments are performed on the UCMerced and AID datasets, which show that when combined with a demonstrative CNN, our method can achieve 2.24% higher accuracy than a plain CNN and can obtain an overall accuracy of 92.43% when combined with AlexNet. The results indicate that the proposed method can effectively improve CNN performance using easy-to-access unlabeled images and thus will enhance the performance of land-use scene classification especially when a large-scale labeled dataset is unavailable.

Land image recognition and classification and land environment detection are important research fields in remote sensing applications. Because of the diversity and complexity of different tasks of land environment recognition and classification, it is difficult for researchers to use a single model to achieve the best performance in scene classification of multiple remote sensing land images. Therefore, to determine which model is the best for the current recognition classification tasks, it is often necessary to select and experiment with many different models. However, finding the optimal model is accompanied by an increase in trial-and-error costs and is a waste of researchers’ time, and it is often impossible to find the right model quickly. To address the issue of existing models being too large for easy selection, this paper proposes a multi-path reconfigurable network structure and takes the multi-path reconfigurable residual network (MR-ResNet) model as an example. The reconfigurable neural network model allows researchers to selectively choose the required modules and reassemble them to generate customized models by splitting the trained models and connecting them through modules with different properties. At the same time, by introducing the concept of a multi-path input network, the optimal path is selected by inputting different modules, which shortens the training time of the model and allows researchers to easily find the network model suitable for the current application scenario. A lot of training data, computational resources, and model parameter experience are saved. Three public datasets, NWPU-RESISC45, RSSCN7, and SIRI-WHU datasets, were used for the experiments. The experimental results demonstrate that the proposed model surpasses the classic residual network (ResNet) in terms of both parameters and performance.

We present Self-Guidance, a simple way to train deep neural networks via knowledge distillation. The basic idea is to train sub-network to match the prediction of the full network, so-called "Self-Guidance". Under the "teacher-student" framework, we construct both teacher and student within the same target network. Student network is the sub-networks that randomly skip some portions of the full network. The teacher network is the full network, can be considered as the ensemble of all possible student networks. The training process is performed in a closed-loop: (1) Forward prediction contains two passes that generate student and teacher predictions. (2) Backward distillation allows knowledge transfer from the teacher back to students. Comprehensive evaluations show that our approach improves the generalization ability of deep neural networks to a significant margin. The results prove our superior performance in both image classification on CIFAR10, CIFAR100, and facial expression recognition on FER-2013 and RAF.

Scene classification plays an important role in the interpretation of remotely sensed high-resolution imagery. However, the performance of scene classification strongly relies on the discriminative power of feature representation, which is generally hand-engineered and requires a huge amount of domain-expert knowledge as well as time-consuming hand tuning. Recently, unsupervised feature learning (UFL) provides an alternative way to automatically learn discriminative feature representation from images. However, the performances achieved by conventional UFL methods are not comparable to the state-of-the-art, mainly due to the neglect of locally substantial image structures. This paper presents an improved UFL algorithm based on spectral clustering, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">named</i> UFL-SC, which cannot only adaptively learn good local feature representations but also discover intrinsic structures of local image patches. In contrast to the standard UFL pipeline, UFL-SC first maps the original image patches into a low-dimensional and intrinsic feature space by linear manifold analysis techniques, and then learns a dictionary (e.g., using K-means clustering) on the patch manifold for feature encoding. To generate a feature representation for each local patch, an explicit parameterized feature encoding method, i.e., triangle encoding, is applied with the learned dictionary on the same patch manifold. The holistic feature representation of image scenes is finally obtained by building a bag-of-visual-words (BOW) model of the encoded local features. Experiments demonstrate that the proposed UFL-SC algorithm can extract efficient local features for image scenes and show comparable performance to the state-of-the-art approach on open scene classification benchmark.

In high spatial resolution (HSR) imagery scene classification, it is a challenging task to recognize the high-level semantics from a large volume of complex HSR images. The probabilistic topic model (PTM), which focuses on modeling topics, has been proposed to bridge the so-called semantic gap. Conventional PTMs usually model the images with a dense semantic representation and, in general, one topic space is generated for all the different features. However, this approach fails to consider the sparsity of the semantic representation, the classification quality, as well as the time consumption. In this paper, to solve the above problems, a fully sparse semantic topic model (FSSTM) framework is proposed for HSR imagery scene classification. FSSTM, with an elaborately designed modeling procedure, is able to represent the image with sparse but representative semantics. Based on this framework, the topic weights of multiple features are exploited by solving a concave maximization problem, which improves the fusion of the discriminative semantic information at the topic level. Meanwhile, the sparsity and representativeness of the topics generated by FSSTM guarantee that the image is adaptive to the change of a topic number. FSSTM can consistently achieve a good performance with a limited number of training samples, and is robust for HSR image scene classification. The experimental results obtained with three different types of HSR image data sets confirm that the proposed algorithm is effective in improving the performance of scene classification, and is highly efficient in discovering the semantics of HSR images when compared with the state-of-the-art PTM methods.

In the recent decades, various techniques based on deep convolutional neural networks (DCNNs) have been applied to scene classification. Most of the techniques are established upon single-spectral images such that environmental conditions may greatly affect the quality of images in the visible (RGB) spectrum. One remedy for this downside is to merge the infrared (IR) with the visible spectrum for gaining the complementary information in comparison with the unimodal analysis. This paper incorporates the RGB, IR and near-infrared (NIR) images into a multispectral analysis for scene classification. For this purpose, two strategies are adopted. In the first strategy, each RGB, IR and NIR image is separately applied to DCNNs and then classified according to the output score of each network. In addition, an optimal decision threshold is obtained based on the same output score of each network. In the second strategy, three image components are extracted from each type of image using wavelet transform decomposition. Independent DCNNs are then trained on the image components of all the scene classes. Eventually, the final classification of the scene is accomplished through an appropriate ensemble architecture. The use of this architecture alongside a transfer learning approach and simple classifiers leads to lesser computational costs in small datasets. These experiments reveal the superiority of the proposed method over the state-of-the-art architectures in terms of the accuracy of scene classification.

Common acoustic properties that different classes share degrades the performance of acoustic scene classification systems. This results in a phenomenon where a few confusing pairs of acoustic scenes dominate a significant proportion of all misclassified audio segments. In this article, we propose adopting a knowledge distillation framework that trains deep neural networks using soft labels. Soft labels, extracted from another pre-trained deep neural network, are used to reflect the similarity between different classes that share similar acoustic properties. We also propose utilizing specialist models to provide additional soft labels. Each specialist model in this study refers to a deep neural network that concentrates on discriminating a single pair of acoustic scenes that are frequently misclassified. Self multi-head attention is explored for training specialist deep neural networks to further concentrate on target pairs of classes. The goal of this article is to train a single deep neural network that demonstrates performance equivalent to, or higher than, an ensemble of multiple models, by distilling the knowledge from several models. Diverse experiments conducted using the detection and classification of acoustic scenes and events 2019 task 1-a dataset demonstrate that the knowledge distillation framework is effective in acoustic scene classification. Specialist models successfully decrease the number of misclassified audio segments in the target classes. The final single model with the proposed method that is trained by the proposed knowledge distillation from several models, including specialists trained using an attention mechanism, shows a classification accuracy of 77.63 %, higher than an ensemble of the baseline and multiple specialists.

Aerial scene classification is an active and challenging problem in high-resolution remote sensing imagery understanding. Deep learning models, especially convolutional neural networks (CNNs), have achieved prominent performance in this field. The extraction of deep features from the layers of a CNN model is widely used in these CNN-based methods. Although the CNN-based approaches have obtained great success, there is still plenty of room to further increase the classification accuracy. As a matter of fact, the fusion with other features has great potential for leading to the better performance of aerial scene classification. Therefore, we propose two effective architectures based on the idea of feature-level fusion. The first architecture, i.e., texture coded two-stream deep architecture, uses the raw RGB network stream and the mapped local binary patterns (LBP) coded network stream to extract two different sets of features and fuses them using a novel deep feature fusion model. In the second architecture, i.e., saliency coded two-stream deep architecture, we employ the saliency coded network stream as the second stream and fuse it with the raw RGB network stream using the same feature fusion model. For sake of validation and comparison, our proposed architectures are evaluated via comprehensive experiments with three publicly available remote sensing scene datasets. The classification accuracies of saliency coded two-stream architecture with our feature fusion model achieve 97.79%, 98.90%, 94.09%, 95.99%, 85.02%, and 87.01% on the UC-Merced dataset (50% and 80% training samples), the Aerial Image Dataset (AID) (20% and 50% training samples), and the NWPU-RESISC45 dataset (10% and 20% training samples), respectively, overwhelming state-of-the-art methods.

동일한 장소에서도 매우 다양한 음향이 발생하고, 서로 다른 장소에서도 유사한 음향이 발생하기 때문에 훈련 데이터가 적거나, 훈련 단계에서 일부 음향만 고려된 경우 음향 상황 인지 성능을 보장할 수 없다. 이러한 문제점을 해결하기 위한 방법으로 Bag of Words (BOW) 기반 히스토그램 특징이 소개되었다. 하지만 BOW 기반 히스토그램 특징은 일정 시간동안 발생한 음향의 분포를 이용하기 때문에 음향이 발생한 순차적인 정보는 고려할 수 없다. 음향 상황 인지에서 일정 시간 동안 발생한 음향의 주기성과 지속성은 상황을 인지하는데 중요한 정보가 될 수 있다. 따라서 본 논문에서는 재발량 분석을 이용하여 주기성과 지속성에 대한 특징을 추출하였다. 인식 실험에서 재발량 분석을 통해 추출된 특징을 함께 사용한 경우 기존 방법들 보다 향상된 성능을 확인했다. Since a variety of sound occur in same place and similar sound occurs in other places, the performance of acoustic scene classification is not guaranteed in case of insufficient training data. A Bag of Words (BOW) based histogram feature is foreseen as a method to overcome the problem. However, since the histogram features is made by using a feature distribution, the ordering of sequence of features is ignored. A temporal information such as periodicity and stationarity are also important for acoustic scene classification. In this paper, temporal features about a periodicity and a stationarity are extracted by using a recurrent quantification analysis. In the experiment, performance of the proposed method is shown better than other baseline methods.

In the human brain, the incoming light to the retina is transformed into meaningful representations that allow us to interact with the world. In a similar vein, the RGB pixel values are transformed by a deep neural network (DNN) into meaningful representations relevant to solving a computer vision task it was trained for. Therefore, in my research, I aim to reveal insights into the visual representations in the human visual cortex and DNNs solving vision tasks. In the previous decade, DNNs have emerged as the state-of-the-art models for predicting neural responses in the human and monkey visual cortex. Research has shown that training on a task related to a brain region’s function leads to better predictivity than a randomly initialized network. Based on this observation, we proposed that we can use DNNs trained on different computer vision tasks to identify functional mapping of the human visual cortex. To validate our proposed idea, we first investigate a brain region occipital place area (OPA) using DNNs trained on scene parsing task and scene classification task. From the previous investigations about OPA’s functions, we knew that it encodes navigational affordances that require spatial information about the scene. Therefore, we hypothesized that OPA’s representation should be closer to a scene parsing model than a scene classification model as the scene parsing task explicitly requires spatial information about the scene. Our results showed that scene parsing models had representation closer to OPA than scene classification models thus validating our approach. We then selected multiple DNNs performing a wide range of computer vision tasks ranging from low-level tasks such as edge detection, 3D tasks such as surface normals, and semantic tasks such as semantic segmentation. We compared the representations of these DNNs with all the regions in the visual cortex, thus revealing the functional representations of different regions of the visual cortex. Our results highly converged with previous investigations of these brain regions validating the feasibility of the proposed approach in finding functional representations of the human brain. Our results also provided new insights into underinvestigated brain regions that can serve as starting hypotheses and promote further investigation into those brain regions. We applied the same approach to find representational insights about the DNNs. A DNN usually consists of multiple layers with each layer performing a computation leading to the final layer that performs prediction for a given task. Training on different tasks could lead to very different representations. Therefore, we first investigate at which stage does the representation in DNNs trained on different tasks starts to differ. We further investigate if the DNNs trained on similar tasks lead to similar representations and on dissimilar tasks lead to more dissimilar representations. We selected the same set of DNNs used in the previous work that were trained on the Taskonomy dataset on a diverse range of 2D, 3D and semantic tasks. Then, given a DNN trained on a particular task, we compared the representation of multiple layers to corresponding layers in other DNNs. From this analysis, we aimed to reveal where in the network architecture task-specific representation is prominent. We found that task specificity increases as we go deeper into the DNN architecture and similar tasks start to cluster in groups. We found that the grouping we found using representational similarity was highly correlated with grouping based on transfer learning thus creating an interesting application of the approach to model selection in transfer learning. During previous works, several new measures were introduced to compare DNN representations. So, we identified the commonalities in different measures and unified different measures into a single framework referred to as duality diagram similarity. This work opens up new possibilities for similarity measures to understand DNN representations. While demonstrating a much higher correlation with transfer learning than previous state-of-the-art measures we extend it to understanding layer-wise representations of models trained on the Imagenet and Places dataset using different tasks and demonstrate its applicability to layer selection for transfer learning. In all the previous works, we used the task-specific DNN representations to understand the representations in the human visual cortex and other DNNs. We were able to interpret our findings in terms of computer vision tasks such as edge detection, semantic segmentation, depth estimation, etc. however we were not able to map the representations to human interpretable concepts. Therefore in our most recent work, we developed a new method that associates individual artificial neurons with human interpretable concepts. Overall, the works in this thesis revealed new insights into the representation of the visual cortex and DNNs...