Robust Convolutional Neural Network Research Articles

Hierarchical and Robust Convolutional Neural Network for Very High-Resolution Remote Sensing Object Detection

Object detection is a basic issue of very high-resolution remote sensing images (RSIs) for automatically labeling objects. At present, deep learning has gradually gained the competitive advantage for remote sensing object detection, especially based on convolutional neural networks (CNNs). Most of the existing methods use the global information in the fully connected feature vector and ignore the local information in the convolutional feature cubes. However, the local information can provide spatial information, which is helpful for accurate localization. In addition, there are variable factors, such as rotation and scaling, which affect the object detection accuracy in RSIs. In order to solve these problems, this paper presents a hierarchical robust CNN. First, multiscale convolutional features are extracted to represent the hierarchical spatial semantic information. Second, multiple fully connected layer features are stacked together so as to improve the rotation and scaling robustness. Experiments on two data sets have shown the effectiveness of our method. In addition, a large-scale high-resolution remote sensing object detection data set is established to make up for the current situation that the existing data set is insufficient or too small. The data set is available at https://github.com/CrazyStoneonRoad/TGRS-HRRSD-Dataset .

Implementation of Improved Ship-Iceberg Classifier Using Deep Learning

Abstract The application of synthetic aperture radar (SAR) for ship and iceberg monitoring is important to carry out marine activities safely. The task of differentiating the two target classes, i.e. ship and iceberg, presents a challenge for operational scenarios. The dataset comprising SAR images of ship and iceberg poses a major challenge, as we are provided with a small number of labeled samples in the training set compared to a large number of unlabeled test samples. This paper proposes a semisupervised learning approach known as pseudolabeling to deal with the insufficient amount of training data. By adopting this approach, we make use of both labeled data (supervised learning) and unlabeled data (unsupervised learning) to build a robust convolutional neural network model that results in a superior binary classification performance of the proposed method.

Relative coordinates constraint for face alignment

We present a practical approach to improve the precision of face alignment for a single image. Recently, face alignment is deemed as a regression problem, and convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are utilized to predict the coordinates of facial landmarks. However, most existing methods only adopt Euclidean loss as the optimization target for each landmark, and neglect the correlations between them, which we think may be inappropriate. To address this issue, in this paper, we introduce a novel Relative Coordinates Constraint (RCC) loss function for face alignment, which considers the relative coordinates between any pairs of landmarks as a new supervision signal. More importantly, we prove that the proposed RCC loss function is trainable and can be easily incorporated in existing CNNs optimization procedure. With the joint supervision of Euclidean loss and RCC loss, we train a robust and light CNNs framework for face alignment. Extensive experimental results on several datasets show that the precision of face alignment improved significantly by the proposed RCC loss and quantitative results are comparable to state-of-the-art methods (mean error 5.39 on 300-W and 6.99 on AFLW). In addition, the proposed framework is also an efficient solution (300 FPS on CPU). We share the implementation code of our proposed methods at https://github.com/nianfudong/RCC-loss.

Unsupervised Seismic Random Noise Attenuation Based on Deep Convolutional Neural Network

Random noise attenuation is one of the most essential steps in seismic signal processing. We propose a novel approach to attenuate seismic random noise based on deep convolutional neural network (CNN) in an unsupervised learning manner. First, normalization and patch sampling are required to build training dataset and test dataset from raw noisy data. Instead of using synthetic noise-free data or denoised results via conventional methods as training labels, we adopt only the training set constructed from the raw noisy data as the input and design a robust deep CNN that just relies on the noisy input to learn the hidden features. The cross-entropy is chosen as the error criterion for establishing the cost function, which is minimized by the back-propagation algorithm to obtain the optimized parameters of the network. Then, we can reconstruct all patches of the test dataset via the optimized CNN. After patching processing and inverse normalization, the final denoised result can be obtained from reconstructed patches. Experimental tests on synthetic and real data demonstrate the effectiveness and superiority of the proposed method compared with state-of-the-art denoising methods.

New Machine Learning Developments in ROOT/TMVA

The Toolkit for Multivariate Analysis, TMVA, the machine learning package integrated into the ROOT data analysis framework, has recently seen improvements to its deep learning module, parallelisation of multivariate methods and cross validation. Performance benchmarks on datasets from high-energy physics are presented with a particular focus on the new deep learning module which contains robust fully-connected, convolutional and recurrent deep neural networks implemented on CPU and GPU architectures. Both dense and convo-lutional layers are shown to be competitive on small-scale networks suitable for high-level physics analyses in both training and in single-event evaluation. Par-allelisation efforts show an asymptotical 3-fold reduction in boosted decision tree training time while the cross validation implementation shows significant speed up with parallel fold evaluation.

Coffee Grading with Convolutional Neural Networks using Small Datasets with High Variance

Convolutional Neural Networks (CNNs) have been established as a powerful class of models for image recognition problems. Despite their success in other areas, CNNs have been applied only for very limited agricultural applications due to the need for large datasets. The aim of this research is to design a robust CNN model that classifies raw coffee beans into their 12 quality grades using small datasets which have high data variability. The dataset contains images of raw coffee beans acquired in two sets using different acquisition technique under varying illuminations which poses a complex challenge to designing a robust model. To design the model, preprocessing techniques were applied to the input in order to reduce task irrelevant features. But adding the preprocessing techniques did not improve the performance of the CNN model for our dataset. We have also used ensemble methods to solve the high variance that exists in networks when working with small datasets. Finally, we were able to design a model that classifies the beans into their quality grades with an accuracy of 89.01% on the test dataset.

Data Augmentation Based on Attributed Scattering Centers to Train Robust CNN for SAR ATR

Driven by the good classification performance of the convolutional neural network (CNN), this study proposes a CNN-based synthetic aperture radar (SAR) target recognition method. In this paper, a novel data augmentation algorithm is proposed via target reconstruction based on attributed scattering centers (ASC). The ASCs reflect the electromagnetic phenomenon of SAR targets, which can be used to reconstruct the target’s characteristics. The sparse representation (SR) algorithm is first employed to extract the ASCs from a single SAR image. Afterward, some of the extracted ASCs are selected to reconstruct the target’s image. By repeating the process, many new images can be generated as available training samples. In the classification stage, a CNN architecture is designed and trained by the augmented samples. For the test sample, it is also reconstructed using all the extracted ASCs thus relieving the interferences caused by the clutters or noises in the background. Finally, the reconstructed image from the test sample is classified based on the trained CNN. The reconstructed image from the ASCs can reduce the clutters and noises thus enhancing the image quality. More importantly, the generated new training samples could cover more operating conditions, which may probably occur in SAR target recognition. Therefore, the trained CNN can work more robustly in different situations. In the experiments, the moving and stationary target acquisition and recognition (MSTAR) dataset is used to evaluate the performance of the proposed approach. This method could classify the 10 classes of targets with an accuracy of 99.48% under the standard operating condition (SOC). For the extended operating conditions like configuration variance, depression angle variance, noise corruption, and partial occlusion, the proposed method also displays superior performance over some baseline algorithms drawn from state-of-the-art literature.

A scale robust convolutional neural network for automatic building extraction from aerial and satellite imagery

ABSTRACTIdentifying buildings from remote sensing imagery has been a challenge due to uncertainties from remote sensing imagery and variations in building structure and texture. In this study, we develop a scale robust CNN structure to improve the segmentation accuracy of building data from high-resolution aerial and satellite images. Based on a fully convolutional network, we introduce two Atrous convolutions on the first two lowest-scale layers, respectively, in the decoding step, aiming at enlarging the sight-of-view and integrate semantic information of large buildings. Then, a multi-scale aggregation strategy is applied. The last feature maps of each scale are used to predict the corresponding building labels, and further up-sampled to the original scale and concatenated for the final prediction. In addition, we introduce a combined data augmentation and relative radiometric calibration method for multi-source building extraction. The method enlarges sample spaces and hence the generalization ability of the deep learning models. We validate our developed methods with an aerial dataset of more than 180, 000 buildings with various architectural types, and a satellite image dataset consists of more than 29,000 buildings. The results are compared with several most recent studies. The comparison result shows our neural network outperformed other studies, especially in segmenting scenes of large buildings. The test on transfer learning from aerial dataset to satellite dataset showed our augmentation strategy significantly improved the prediction accuracy; however, further studies are needed to improve the generalization ability of the CNN model.

Traffic sign recognition using weighted multi‐convolutional neural network

Traffic signs play a crucial role in regulating traffic and facilitating cautious driving. Automatic traffic sign recognition is one of the key tasks in autonomous driving. Accuracy in the classification of traffic signs is therefore very important for the navigation of a vehicle. Here, a reliable and robust convolutional neural network (CNN) is presented for classifying these signs. The proposed classifier is a weighted multi-CNN trained with a novel methodology. It achieves a near state-of-the-art recognition rate of 99.59% when tested on the German traffic sign recognition benchmark dataset. Compared to the existing classifiers, the proposed one is a low-complexity network that recognises a test image in 10 ms when running on an NVIDIA 980 Ti GPU system. The results demonstrate its suitability and reliability in high-speed driving scenarios.

A robust convolutional neural network for online smartphone-based human activity recognition

The online smartphone-based human activity recognition (HAR) has a variety of applications such as fitness tracking, healthcare… etc. Currently, the signals generated from smartphone-embedded sensors are used for HAR systems. The smartphone-embedded sensors are utilized in order to provide an unobtrusive platform for HAR. In this paper, we propose a deep convolution neural network (CNN) model that provides an effective and efficient smartphone-based HAR system. For automatic local features extraction from the raw time-series data, we use the CNN while simple time-domain statistical features are used to extract more distinguishable features. Furthermore, we explore the impact of a novel data augmentation on the recognition accuracy of the proposed model. The performance of the proposed method is evaluated using two public data sets (UCI and WISDM) which are collected using smartphones

Overfitting remedy by sparsifying regularization on fully-connected layers of CNNs

Abstract Deep learning, especially Convolutional Neural Networks (CNNs), has been widely applied in many domains. The large number of parameters in a CNN allow it to learn complex features, however, they may tend to hinder generalization by over-fitting training data. Despite many previously proposed regularization methods, over-fitting is still a problem in training a robust CNN. Among many factors that lead to over-fitting, the numerous parameters of fully-connected layers (FCLs) of a typical CNN should be taken into account. This paper proposes the SparseConnect, a simple idea which alleviates over-fitting by sparsifying connections to FCLs. Experimental results on three benchmark datasets MNIST, CIFAR10 and ImageNet show that the SparseConnect outperforms several state-of-the-art regularization methods.

Automated classification of osteomeatal complex inflammation on computed tomography using convolutional neural networks.

Convolutional neural networks (CNNs) are advanced artificial intelligence algorithms well suited to image classification tasks with variable features. These have been used to great effect in various real-world applications including handwriting recognition, face detection, image search, and fraud prevention. We sought to retrain a robust CNN with coronal computed tomography (CT) images to classify osteomeatal complex (OMC) occlusion and assess the performance of this technology with rhinologic data. The Google Inception-V3 CNN trained with 1.28 million images was used as the base model. Preoperative coronal sections through the OMC were obtained from 239 patients enrolled in 2 prospective chronic rhinosinusitis (CRS) outcomes studies, labeled according to OMC status, and mirrored to obtain a set of 956 images. Using this data, the classification layer of Inception-V3 was retrained in Python using a transfer learning method to adapt the CNN to the task of interpreting sinonasal CT images. The retrained neural network achieved 85% classification accuracy for OMC occlusion, with a 95% confidence interval for algorithm accuracy of 78% to 92%. Receiver operating characteristic (ROC) curve analysis on the test set confirmed good classification ability of the CNN with an area under the ROC curve (AUC) of 0.87, significantly different than both random guessing and a dominant classifier that predicts the most common class (p < 0.0001). Current state-of-the-art CNNs may be able to learn clinically relevant information from 2-dimensional sinonasal CT images with minimal supervision. Future work will extend this approach to 3-dimensional images in order to further refine this technology for possible clinical applications.

A Novel Deep Learning Approach for Recognizing Stereotypical Motor Movements within and across Subjects on the Autism Spectrum Disorder.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by persistent difficulties including repetitive patterns of behavior known as stereotypical motor movements (SMM). So far, several techniques have been implemented to track and identify SMMs. In this context, we propose a deep learning approach for SMM recognition, namely, convolutional neural networks (CNN) in time and frequency-domains. To solve the intrasubject SMM variability, we propose a robust CNN model for SMM detection within subjects, whose parameters are set according to a proper analysis of SMM signals, thereby outperforming state-of-the-art SMM classification works. And, to solve the intersubject variability, we propose a global, fast, and light-weight framework for SMM detection across subjects which combines a knowledge transfer technique with an SVM classifier, therefore resolving the “real-life” medical issue associated with the lack of supervised SMMs per testing subject in particular. We further show that applying transfer learning across domains instead of transfer learning within the same domain also generalizes to the SMM target domain, thus alleviating the problem of the lack of supervised SMMs in general.

Classification Of Skin Lesions Using An Ensemble Of Deep Neural Networks.

Skin cancer is among the deadliest variants of cancer if not recognized and treated in time. This work focuses on the identification of this disease using an ensemble of state-of-the-art deep learning approaches. More specifically, we propose the aggregation of robust convolutional neural networks (CNNs) into one neural net architecture, where the final classification is achieved based on the weighted output of the member CNNs. Since our framework is realized within a single neural net architecture, all the parameters of the member CNNs and the weights applied in the fusion can be determined by backpropagation routinely applied for such tasks. The presented ensemble consists of the CNNs AlexNet, VGGNet, GoogLeNet, all of which have been won in subsequent years the most prominent worldwide image classification challenge ImageNet. For an objective evaluation of our approach, we have tested its performance on the official test database of the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 challenge on Skin Lesion Analysis Towards Melanoma Detection dedicated to skin cancer recognition. Our experimental studies show that the proposed approach is competitive in this field. Moreover, the ensemble-based approach outperformed all of its member CNNs.

Flower classification using deep convolutional neural networks

Flower classification is a challenging task due to the wide range of flower species, which have a similar shape, appearance or surrounding objects such as leaves and grass. In this study, the authors propose a novel two‐step deep learning classifier to distinguish flowers of a wide range of species. First, the flower region is automatically segmented to allow localisation of the minimum bounding box around it. The proposed flower segmentation approach is modelled as a binary classifier in a fully convolutional network framework. Second, they build a robust convolutional neural network classifier to distinguish the different flower types. They propose novel steps during the training stage to ensure robust, accurate and real‐time classification. They evaluate their method on three well known flower datasets. Their classification results exceed 97% on all datasets, which are better than the state‐of‐the‐art in this domain.

Ear verification under uncontrolled conditions with convolutional neural networks

The capabilities of biometric systems have recently made extraordinary leaps by the emergence of deep learning. However, due to the lack of enough training data, the applications of the deep neural network in the ear recognition filed have run into the bottleneck. Moreover, the effect of fine-tuning from some pre-trained models is far less than expected due to the diversity among different tasks. Therefore, the authors propose a large-scale ear database and explore the robust convolutional neural network (CNN) architecture for the ear feature representation. The images in this USTB-Helloear database were taken under uncontrolled conditions with illumination, pose variation and different level of ear occlusions. Then they fine-tuned and modified some deep models on the proposed database through the ear verification experiments. First, they replaced the last pooling layers by spatial pyramid pooling layers to fit arbitrary data size and obtain multi-level features. In the training phase, the CNNs were trained both under the supervision of the softmax loss and centre loss to obtain more compact and discriminative features to identify unseen ears. Finally, three CNNs with different scales of ear images were assembled as the multi-scale ear representations for ear verification. The experimental results demonstrate the effectiveness of the proposed modified CNN deep model.

A Rank Decomposed Statistical Error Compensation Technique for Robust Convolutional Neural Networks in the Near Threshold Voltage Regime

There has been a growing interest in implementing complex machine learning algorithms such as convolutional neural networks (CNNs) on lower power embedded platforms to enable on-device learning and inference. Many of these platforms are to be deployed as low power sensor nodes with low to medium throughput requirement. Near threshold voltage (NTV) designs are well-suited for these applications but suffer from a significant increase in variations. In this paper, we propose a variation-tolerant architecture for CNNs capable of operating in NTV regime for energy efficiency. A statistical error compensation (SEC) technique referred to as rank decomposed SEC (RD-SEC) is proposed. The key idea is to exploit inherent redundancy within matrix-vector multiplication (or dot product ensemble), a power-hungry operation in CNNs, to derive low-cost estimators for error detection and compensation. When evaluated in CNNs for both the MNIST and CIFAR-10 datasets, simulation results in 45 nm CMOS show that RD-SEC enables robust CNNs operating in the NTV regime. Specifically, the proposed architecture can achieve up to 11 × improvement in variation tolerance and enable up to 113 × reduction in the standard deviation of detection accuracy P d e t while incurring marginal degradation in the median detection accuracy.

Voice Pathology Detection Using Deep Learning on Mobile Healthcare Framework

The feasibility and popularity of mobile healthcare are currently increasing. The advancement of modern technologies, such as wireless communication, data processing, the Internet of Things, cloud, and edge computing, makes mobile healthcare simpler than before. In addition, the deep learning approach brings a revolution in the machine learning domain. In this paper, we investigate a voice pathology detection system using deep learning on the mobile healthcare framework. A mobile multimedia healthcare framework is also designed. In the voice pathology detection system, voices are captured using smart mobile devices. Voice signals are processed before being fed to a convolutional neural network (CNN). We use a transfer learning technique to use the existing robust CNN models. In particular, the VGG-16 and CaffeNet models are investigated in the paper. The Saarbrucken voice disorder database is used in the experiments. Experimental results show that the voice pathology detection accuracy reaches up to 97.5% using the transfer learning of CNN models.

Video Satellite Imagery Super Resolution via Convolutional Neural Networks

Video satellite imagery is a new technique for earth dynamic observation and has a wide range of uses in environmental fields. Despite its capability of dynamic targets’ detection, it sustains a serious restriction of the image quality due to the degradation and compression in its imaging process. Hence, the super-resolution (SR) reconstruction on these compressed low-spatial-resolution images is of significance to afterward ground objects recognition and detection tasks. Based on the recent proposed state-of-the-art convolutional neural networks (CNNs) SR methods, we proposed an SR method which could get more precise reconstructed high-spatial-resolution images. Trained with Gaofen-2 satellite images, a robust CNN model specified in satellite image SR is obtained. Experimentally, the reconstruction results on Jilin-1 mission satellite images validate the effectiveness of our method.

SparseConnect: regularising CNNs on fully connected layers

Deep convolutional neural networks (CNNs) have achieved unprecedented success in many domains. The numerous parameters allow CNNs to learn complex features, but also tend to hinder generalisation by over-fitting training data. Despite many previously proposed regularisation methods, over-fitting is one of many problems in training a robust CNN. Among many factors that may lead to over-fitting, the numerous parameters of fully connected layers (FCLs) of a typical CNN are a contributor to the over-fitting problem. The authors propose SparseConnect, which alleviates over-fitting by sparsifying connections to FCLs. Experimental results on three benchmark datasets MNIST and CIFAR10 show SparseConnect outperforms several state-of-the-art regularisation methods.