Deep Classification Research Articles

Robust Adversarial Attack Against Explainable Deep Classification Models Based on Adversarial Images With Different Patch Sizes and Perturbation Ratios

In recent years, adversarial attack methods have been deceived rather easily on deep neural networks (DNNs). In practice, adversarial patches cause misclassification that can be extremely effective. However, many existing adversarial patches are used for attacking DNNs, and only a few of them apply to both the DNN and its explanation model. In this paper, we present different adversarial patches that misguide the prediction of DNN models and change the cause of prediction results of interpretation models, such as gradient-weighted class activation mapping. The proposed adversarial patches have appropriate location and perturbation ratios, which comprise visible or less visible adversarial patches. In addition, image patches within small arrays are localized without covering or overlapping with any of the main objects in a natural image. In particular, we generate two adversarial patches that cover only 3% and 1.5% of the pixels in the original image, while they do not cover the main objects in the natural image. Our experiments are performed using four pre-trained DNN models and the ImageNet dataset. We also examine the inaccurate results of the interpretation models through mask and heatmap visualization. The proposed adversarial attack method could be a reference for developing robust network interpretation models that are more reliable for the decision-making process of pre-trained DNN models.

Hybrid Pre-Training Strategy for Deep Denoising Neural Networks and Its Application in Machine Fault Diagnosis

Removing or at least attenuating the effect of noise is of significance for machine fault diagnosis under noisy measuring environments. In this paper, deep denoising neural networks are developed to decrease the influence of noise on the modeling accuracy of deep neural networks (DNN). Among the network, a reproducing kernel Hilbert space (RKHS) based denoising layer is designed as the first layer to reduce noise. However, due to the unstable gradients problem of the earlier layers in DNN, weights to-be-optimized in the first denoising layer are tough to update. In addition, the overfitting problem caused by multiple layers also affects the generalization ability of the model. To overcome these dilemmas, a novel hybrid pre-training strategy for deep denoising neural networks is proposed. In this strategy, motivated by the greedy algorithm, a shallow supervised network is firstly adopted to pre-train the weight of the denoising layer for alleviating the learning difficulty of the first layers in deep networks. Secondly, feature extraction layers with a large number of layers are pre-trained by an autoencoder-based unsupervised network. After that, weights of the denoising and feature extraction layers are transferred to a deep classification network, and the entire network is finally fine-tuned via supervised learning. The hybrid pre-training strategy for deep denoising neural networks is applied to machine fault diagnosis. Experiment studies verify the effectiveness of the proposed hybrid pre-training strategy in intelligent fault diagnosis of machinery under noisy measuring environments.

Incremental Cross-Domain Adaptation for Robust Retinopathy Screening via Bayesian Deep Learning

Retinopathy represents a group of retinal diseases that, if not treated timely, can cause severe visual impairments or even blindness. Many researchers have developed autonomous systems to recognize retinopathy via fundus and optical coherence tomography (OCT) imagery. However, most of these frameworks employ conventional transfer learning and fine-tuning approaches, requiring a decent amount of well-annotated training data to produce accurate diagnostic performance. This paper presents a novel incremental cross-domain adaptation instrument that allows any deep classification model to progressively learn abnormal retinal pathologies in OCT and fundus imagery via few-shot training. Furthermore, unlike its competitors, the proposed instrument is driven via a Bayesian multi-objective function that not only enforces the candidate classification network to retain its prior learned knowledge during incremental training but also ensures that the network understands the structural and semantic relationships between previously learned pathologies and newly added disease categories to effectively recognize them at the inference stage. The proposed framework, evaluated on six public datasets acquired with three different scanners to screen thirteen retinal pathologies, outperforms the state-of-the-art competitors by achieving an overall accuracy and F1 score of 0.9826 and 0.9846, respectively.

Hybrid classification of diffuse liver diseases in ultrasound images using deep convolutional neural networks

Despite the fact that liver biopsy is considered to be the gold standard for detecting diffuse liver diseases, it is an invasive method with numerous side effects. Diffuse liver diagnosis using ultrasound imaging may be influenced by Physician subjectivity. Therefore, an accurate classification of liver diseases remains a notable demand. In this study, to categorize the liver status, a novel deep classifier, comprised of pre-trained deep convolutional neural networks (CNNs) is proposed. Several networks, namely ResNeXt, ResNet18, ResNet34, ResNet50, and AlexNet which concatenated with fully connected networks (FCNs) are used. Extracted deep features using transfer learning can provide sufficient classification information. An FCN can then put images into different states of the disease, namely normal liver, liver hepatitis, and cirrhosis. Two-class (normal/cirrhosis, normal/hepatitis, and cirrhosis/hepatitis) and three-class (normal/cirrhosis/hepatitis) classifiers were trained to distinguish these liver images. Since two-class classifiers showed better performance compared to the three-class classifiers, a hybrid classifier is proposed so as to integrate the weighted probabilities of the classes obtained by means of each individual classifier. Then, a majority voting strategy is employed to select the class with a higher score. The experimental results show an accuracy of 86.4% using ResNet50 with a hybrid classifier for liver images which were classified into three classes. In the distinction between normal and cirrhosis liver as well as normal and hepatitis liver, the results demonstrate the sensitivity and specificity of the first group to be 90.9% and 86.4% and the latter group shows the sensitivity of 90.9%, and specificity of 81.8%.

A novel lifelong learning model based on cross domain knowledge extraction and transfer to classify underwater images

Artificial intelligence based autonomous systems interacting with dynamic environment are required to continuously learn, accumulate and improve the learned knowledge. Currently, most artificial intelligence based systems lack this ability and work in isolated learning paradigm. Human beings follow the continuous learning process by retaining and accumulating the learnt knowledge, and by using the learnt knowledge to solve the problem at hand. In this paper, we present a lifelong learning model, to solve challenging problem of real world underwater image classification. The proposed model is capable to learn from simple problems, accumulates the learnt knowledge by continual learning and uses the learnt knowledge to solve future complex problems of the same or related domain, in a similar way as humans do. In the proposed model, firstly, a deep classification convolutional autoencoder is presented to extract spatially localized features from images by utilizing convolution filters, then a code fragment based learning classifier system, with rich knowledge encoding scheme, is proposed for knowledge representation and transfer. In order to validate the model, experiments are conducted on two underwater images datasets and one in-air images dataset. Experiments results demonstrate that the proposed method outperforms base line method and state-of-the-art convolution neural network (CNN) methods.

Comprehensive Computer-Aided Decision Support Framework to Diagnose Tuberculosis From Chest X-Ray Images: Data Mining Study.

BackgroundTuberculosis (TB) is one of the most infectious diseases that can be fatal. Its early diagnosis and treatment can significantly reduce the mortality rate. In the literature, several computer-aided diagnosis (CAD) tools have been proposed for the efficient diagnosis of TB from chest radiograph (CXR) images. However, the majority of previous studies adopted conventional handcrafted feature-based algorithms. In addition, some recent CAD tools utilized the strength of deep learning methods to further enhance diagnostic performance. Nevertheless, all these existing methods can only classify a given CXR image into binary class (either TB positive or TB negative) without providing further descriptive information.ObjectiveThe main objective of this study is to propose a comprehensive CAD framework for the effective diagnosis of TB by providing visual as well as descriptive information from the previous patients’ database.MethodsTo accomplish our objective, first we propose a fusion-based deep classification network for the CAD decision that exhibits promising performance over the various state-of-the-art methods. Furthermore, a multilevel similarity measure algorithm is devised based on multiscale information fusion to retrieve the best-matched cases from the previous database.ResultsThe performance of the framework was evaluated based on 2 well-known CXR data sets made available by the US National Library of Medicine and the National Institutes of Health. Our classification model exhibited the best diagnostic performance (0.929, 0.937, 0.921, 0.928, and 0.965 for F1 score, average precision, average recall, accuracy, and area under the curve, respectively) and outperforms the performance of various state-of-the-art methods.ConclusionsThis paper presents a comprehensive CAD framework to diagnose TB from CXR images by retrieving the relevant cases and their clinical observations from the previous patients’ database. These retrieval results assist the radiologist in making an effective diagnostic decision related to the current medical condition of a patient. Moreover, the retrieval results can facilitate the radiologists in subjectively validating the CAD decision.

Deep versus conventional machine learning for MRI‐based diagnosis and prediction of Alzheimer’s disease

AbstractBackgroundAccurate diagnosis and prediction of Alzheimer's disease (AD) is important for care planning and clinical trials. Machine learning based on MRI showed promising results for these tasks. While in many medical imaging applications deep neural networks have outperformed conventional approaches, this has not been shown for AD classification. Therefore, this work presents a direct comparison of conventional and deep AD classification.MethodsWe included participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI): 334 AD patients, 520 cognitively normals (CN), 231 mild cognitive impaired (MCI) patients who converted to AD within 3 years (MCIc) and 628 MCI patients who did not convert (MCInc). As input features we considered two MRI‐based images in MNI‐space: raw T1‐weighted images and preprocessed modulated GM maps that encode gray matter density. As conventional classifier, we used a linear support vector machine (SVM). The c‐parameter was optimized with 5‐fold crossvalidation on the training set. The deep classifier was an all convolutional neural network (all‐CNN) that consisted of 7 convolutional blocks and a final softmax classification layer (Fig. 1). A validation set (10%) was used for regularization by early stopping. Classifiers were trained for AD‐CN classification using 20 repetitions of random split (10% test). Confidence intervals (95%CI) on performance measures were computed with the corrected resampled t‐test. Based on AD‐CN results, we selected the best input feature and retrained classifiers on the full AD/CN dataset to evaluate predictive performance (MCIc‐MCInc).ResultsFor AD‐CN classification based on modulated GM maps, the AUC for SVM (0.940; 95%CI: 0.924‐0.955) was similar to that of all‐CNN (0.933; 95%CI: 0.918‐0.948) (Fig. 2). Both with SVM and all‐CNN, classification based on modulated GM maps outperformed classification based on T1w‐images, especially in the case of SVM. Application to MCI conversion prediction (Fig. 3) yielded slightly higher performance for SVM (AUC=0.765) than for all‐CNN (AUC=0.742) (p=0.01 for McNemar’s test).ConclusionDeep and conventional classifiers performed equally well in AD detection. For AD prediction, the conventional classifier yielded slightly higher performance. Both approaches benefitted from the use of modulated GM maps instead of raw T1‐weighted images, while this effect is stronger for the conventional method.

독소 조항 분류를 위한 딥러닝 기반 텍스트 분류 모델

독소 조항 분류를 위한 딥러닝 기반 텍스트 분류 모델

Deep LSAC for Fine-Grained Recognition.

Fine-grained recognition emphasizes the identification of subtle differences among object categories given objects that appear in different shapes and poses. These variances should be reduced for reliable recognition. We propose a fine-grained recognition system that incorporates localization, segmentation, alignment, and classification in a unified deep neural network. The input to the classification module includes functions that enable backward-propagation (BP) in constructing the solver. Our major contribution is to propose a valve linkage function (VLF) for BP chaining and form our deep localization, segmentation, alignment, and classification (LSAC) system. The VLF can adaptively compromise errors of classification and alignment when training the LSAC model. It in turn helps to update the localization and segmentation. We evaluate our framework on two widely used fine-grained object data sets. The performance confirms the effectiveness of our LSAC system.

Adaptive Spider Bird Swarm Algorithm-Based Deep Recurrent Neural Network for Malicious JavaScript Detection Using Box-Cox Transformation

JavaScript is a scripting language that is commonly used in the web pages for providing dynamic functionality in order to enhance user experience. Malicious JavaScript in webpages on internet is an important security issue due to their potentially and universality severe impact. Finding the malicious JavaScript is usually more difficult and time-consuming task in the research community. Hence, an adaptive spider bird swarm algorithm-based deep recurrent neural network (adaptive SBSA-based deep RNN) is proposed for detecting the malicious JavaScript codes in web applications. However, the proposed adaptive SBSA is designed by integrating the adaptive concept with the bird swarm algorithm (BSA) and spider monkey optimization (SMO). With the deep RNN classifier, the complexity issues exists in detecting the malicious codes is effectively resolved through the process of hierarchical computation. Due to the efficiency of the proposed approach, it can evaluate under large real-life datasets.

Review of the State of the Art of Deep Learning for Plant Diseases: A Broad Analysis and Discussion.

Deep learning (DL) represents the golden era in the machine learning (ML) domain, and it has gradually become the leading approach in many fields. It is currently playing a vital role in the early detection and classification of plant diseases. The use of ML techniques in this field is viewed as having brought considerable improvement in cultivation productivity sectors, particularly with the recent emergence of DL, which seems to have increased accuracy levels. Recently, many DL architectures have been implemented accompanying visualisation techniques that are essential for determining symptoms and classifying plant diseases. This review investigates and analyses the most recent methods, developed over three years leading up to 2020, for training, augmentation, feature fusion and extraction, recognising and counting crops, and detecting plant diseases, including how these methods can be harnessed to feed deep classifiers and their effects on classifier accuracy.

A deep learning based HTTP slow DoS classification approach using flow data

The popularity of the Internet introduces many network-enabled services that can be accessed by the user. But the adversaries are trying to deny these critical services to the user through Denial of Service (DoS) attacks. Presently, dealing with DoS attack which targets the application layer using slow traffic rate is one of the key challenges faced by the service providers. In this paper, a deep classification model using flow data is proposed to detect slow DoS attack on HTTP. The classifier is evaluated using CICIDS2017 dataset. The results obtained show that the classifier can obtain 99.61% accuracy.

ArAutoSenti: automatic annotation and new tendencies for sentiment classification of Arabic messages

A corpus-based sentiment analysis approach for messages written in Arabic and its dialects is presented and implemented. The originality of this approach resides in the automation construction of the annotated sentiment corpus, which relies mainly on a sentiment lexicon that is also constructed automatically. For the classification step, shallow and deep classifiers are used with features being extracted applying word embedding models. For the validation of the constructed corpus, we proceed with a manual reviewing and it was found that 85.17% were correctly annotated. This approach is applied on the under-resourced Algerian dialect and the approach is tested on two external test corpora presented in the literature. The obtained results are very encouraging with an F1 score that is up to 88% (on the first test corpus) and up to 81% (on the second test corpus). These results, respectively, represent a 20% and a 6% improvement, respectively, when compared with existing work in the research literature.

A Study on Weeds Retrieval based on Deep Neural Network Classification Model

A Study on Weeds Retrieval based on Deep Neural Network Classification Model

Remote sensing image scene classification using CNN-MLP with data augmentation

Classification of the very high-resolution (VHR) imagery scene has become a challenging problem. The convolutional neural network (CNN) has increased the accuracy in this area due to learning features. However, models based on CNN contain many deep layers for classifying images that are not perfect in describing the relationship between objects within the image. Therefore, an enhanced multilayer perceptron (MLP) depending on Adagrad optimizer is employed in the classification step in this paper as a deep classifier. Motivated by this idea, this paper proposes an effective classification model named CNN-MLP to utilize the benefits of these two techniques: CNN and MLP. The features are generated using pre-trained CNN without fully connected layers. Due to limited training images per class, the proposed model uses data augmentation techniques to expand the training images. Then, an MLP is used to classify the final feature maps into the specified classes. Three public remote sensing datasets of VHR images to evaluate the proposed CNN-MLP model: UC-Merced, Aerial Image (AID), and NWPU-RESISC45 datasets. The experiment's findings show that the proposed method will contribute to higher classification performance relative to state-of-the-art methods.

An Intelligent Diagnosis Method of Brain MRI Tumor Segmentation Using Deep Convolutional Neural Network and SVM Algorithm.

Among the currently proposed brain segmentation methods, brain tumor segmentation methods based on traditional image processing and machine learning are not ideal enough. Therefore, deep learning-based brain segmentation methods are widely used. In the brain tumor segmentation method based on deep learning, the convolutional network model has a good brain segmentation effect. The deep convolutional network model has the problems of a large number of parameters and large loss of information in the encoding and decoding process. This paper proposes a deep convolutional neural network fusion support vector machine algorithm (DCNN-F-SVM). The proposed brain tumor segmentation model is mainly divided into three stages. In the first stage, a deep convolutional neural network is trained to learn the mapping from image space to tumor marker space. In the second stage, the predicted labels obtained from the deep convolutional neural network training are input into the integrated support vector machine classifier together with the test images. In the third stage, a deep convolutional neural network and an integrated support vector machine are connected in series to train a deep classifier. Run each model on the BraTS dataset and the self-made dataset to segment brain tumors. The segmentation results show that the performance of the proposed model is significantly better than the deep convolutional neural network and the integrated SVM classifier.

Towards Understanding and Arguing with Classifiers: Recent Progress

Machine learning and argumentation can potentially greatly benefit from each other. Combining deep classifiers with knowledge expressed in the form of rules and constraints allows one to leverage different forms of abstractions within argumentation mining. Argumentation for machine learning can yield argumentation-based learning methods where the machine and the user argue about the learned model with the common goal of providing results of maximum utility to the user. Unfortunately, both directions are currently rather challenging. For instance, combining deep neural models with logic typically only yields deterministic results, while combining probabilistic models with logic often results in intractable inference. Therefore, we review a novel deep but tractable model for conditional probability distributions that can harness the expressive power of universal function approximators such as neural networks while still maintaining a wide range of tractable inference routines. While this new model has shown appealing performance in classification tasks, humans cannot easily understand the reasons for its decision. Therefore, we also review our recent efforts on how to “argue” with deep models. On synthetic and real data we illustrate how “arguing” with a deep model about its explanations can actually help to revise the model, if it is right for the wrong reasons.

Segmentation of Breast Mass and Diagnosis of Benign and Malignant Breast Tumors Based on Edge Constraint in Pulse Coupled Neural Network

In order to segment breast tumor accurately, an improved Unit-Linking Pulse-Coupled Neural Networks based mammography image segmentation method is proposed. Firstly, the link input and coupled parameter in the original model are improved according to the relationship between this neuron and its neighbors. Then, the improved model is used to segment the breast tumor image to obtain multiple output images. Finally, the gradient algorithm is used to calculate the edges of the original image and each output image respectively, and the minimum mean square error (MMSE) of the two edge images is calculated to find the best output image. The final experimental results indicate that the improved method can accurately segment breast tumor images in different environments. In addition, based on the segmentation results, we use the SVM method to diagnose the type of tumor, and its classification accuracy is much higher than the existing deep classification algorithm.

Deep CNN-based hyperspectral image classification using discriminative multiple spatial-spectral feature fusion

ABSTRACT Convolutional Neural Networks (CNNs) are widely used in various fields, and have shown good performance in hyperspectral image (HSI) classification. Recently, utilizing deep networks to learn spatial-spectral features has become of great interest. However, excessively increasing the depth of network may result in overfitting. Moreover, in HSI classification, the existing network models ignore the strong complementary yet correlated spatial-spectral information among different hierarchical layers. In order to address these two problems, a novel CNN-based method for HSI classification is proposed. Firstly, it considers fusing the outputs of recurrent two layers in each large convolutional block and thereby using the fusion result as the input of next layer, which facilitates the extraction of discriminative features. Next, the spectral-spatial features are extracted by cascading spectral features to four-scale spatial features from shallow to deep layers. Finally, a 11 convolution layer is used to interact and integrate information across channels. Without increasing the number of training samples and the size of pixel patches at the training stage, the proposed approach achieves the state-of-the-art results in the experiment on three well-known hyperspectral images.

Appearance variation adaptation tracker using adversarial network

Visual trackers using deep neural networks have demonstrated favorable performance in object tracking. However, training a deep classification network using overlapped initial target regions may lead an overfitted model. To increase the model generalization, we propose an appearance variation adaptation (AVA) tracker that aligns the feature distributions of target regions over time by learning an adaptation mask in an adversarial network. The proposed adversarial network consists of a generator and a discriminator network that compete with each other over optimizing a discriminator loss in a mini-max optimization problem. Specifically, the discriminator network aims to distinguish recent target regions from earlier ones by minimizing the discriminator loss, while the generator network aims to produce an adaptation mask to maximize the discriminator loss. We incorporate a gradient reverse layer in the adversarial network to solve the aforementioned mini-max optimization in an end-to-end manner. We compare the performance of the proposed AVA tracker with the most recent state-of-the-art trackers by doing extensive experiments on OTB50, OTB100, and VOT2016 tracking benchmarks. Among the compared methods, AVA yields the highest area under curve (AUC) score of 0.712 and the highest average precision score of 0.951 on the OTB50 tracking benchmark. It achieves the second best AUC score of 0.688 and the best precision score of 0.924 on the OTB100 tracking benchmark. AVA also achieves the second best expected average overlap (EAO) score of 0.366, the best failure rate of 0.68, and the second best accuracy of 0.53 on the VOT2016 tracking benchmark.