Convolutional Classifier Research Articles

Defect detection through customised reduction and hybrid convolution classification over super-pixel clusters

Defect detection through customised reduction and hybrid convolution classification over super-pixel clusters

Defect Detection through Customised Reduction and Hybrid Convolution Classification over Super-pixel Clusters

Defect Detection through Customised Reduction and Hybrid Convolution Classification over Super-pixel Clusters

Few-Shot Learning for Radar Signal Recognition Based on Tensor Imprint and Re-Parameterization Multi-Channel Multi-Branch Model

In an ever-increasingly complex electromagnetic environment, automatic radar signal recognition is becoming vital. Convolutional neural networks have been widely used for radar signal recognition, but deep learning-based algorithms only recognize trained classes. Recognizing novel radar signals with few-shot samples in an open environment is still a challenging research problem. In this letter, a few-shot learning algorithm based on the tensor imprint algorithm and convolutional classification layer is proposed for radar signal recognition, and the proposed convolutional classification layer can avoid spatial information loss caused by the global pooling layer and the fully connected layer. In addition, the lightweight re-parameterization multi-channel multi-branch convolutional neural network (RepMCMBNet) is proposed for feature extraction. The model is trained on a dataset containing 8 types of radar signals, and achieves high recognition accuracy in a test dataset containing 12 types of radar signals. The overall recognition accuracy of the proposed tensor imprint algorithm achieves 93.9% at −6 dB when the number of samples is 5.

Quantitative detection of Aflatoxin B1 by subpixel CNN regression

Aflatoxin is a highly toxic substance dispersed in peanuts, which seriously harms the health of humans and animals. In this paper, we propose a new method for aflatoxin B1(AFB1) detection inspired by quantitative remote sensing. Firstly, we obtained the relative content of AFB1 at the sub-pixel level by subpixel decomposition (endmember extraction, nonnegative matrix decomposition). Then we modified the transfer learning models (LeNet5, AlexNet, VGG16, and ResNet18) to construct a deep learning regression network for quantitative detection of AFB1. There are 67,178 pixels used for training and 67,164 pixels used for testing. After subpixel decomposition, each aflatoxin pixel was determined to contain content, and each pixel had 400 hyperspectral values (415–799 nm). The experimental results showed that, among the four models, the modified ResNet18 model achieved the best effect, with R2 of 0.8898, RMSE of 0.0138, and RPD of 2.8851. Here, we implemented a sub-pixel model for quantitative AFB1 detection and proposed a regression method based on deep learning. Meanwhile, the modified convolution classification model has high predictive ability and robustness. This method provides a new scheme in designing the sorting machine and has practical value.

Parkinson\u2019s disease diagnosis using convolutional neural networks and figure-copying tasks

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that causes abnormal movements and an array of other symptoms. An accurate PD diagnosis can be a challenging task as the signs and symptoms, particularly at an early stage, can be similar to other medical conditions or the physiological changes of normal ageing. This work aims to contribute to the PD diagnosis process by using a convolutional neural network, a type of deep neural network architecture, to differentiate between healthy controls and PD patients. Our approach focuses on discovering deviations in patient’s movements with the use of drawing tasks. In addition, this work explores which of two drawing tasks, wire cube or spiral pentagon, are more effective in the discrimination process. With 93.5% accuracy, our convolutional classifier, trained with images of the pentagon drawing task and augmentation techniques, can be used as an objective method to discriminate PD from healthy controls. Our compact model has the potential to be developed into an offline real-time automated single-task diagnostic tool, which can be easily deployed within a clinical setting.

Geometric deep learning on brain shape predicts sex and age.

The complex relationship between the shape and function of the human brain remains elusive despite extensive studies of cortical folding over many decades. The analysis of cortical gyrification presents an opportunity to advance our knowledge about this relationship, and better understand the etiology of a variety of pathologies involving diverse degrees of cortical folding abnormalities. Hypothesis-driven surface-based approaches have been shown to be particularly efficient in their ability to accurately describe unique features of the folded sheet topology of the cortical ribbon. However, the utility of these approaches has been blunted by their reliance on manually defined features aiming to capture the relevant geometric properties of cortical folding. In this paper, we propose an entirely novel, data-driven deep-learning based method to analyze the brain's shape that eliminates this reliance on manual feature definition. This method builds on the emerging field of geometric deep-learning and uses traditional convolutional neural network architecture uniquely adapted to the surface representation of the cortical ribbon. This method is a complete departure from prior brain MRI CNN investigations, all of which have relied on three dimensional MRI data and interpreted features of the MRI signal for prediction. MRI data from 6410 healthy subjects obtained from 11 publicly available data repositories were used for analysis. Ages ranged from 6 to 89 years. Both inner and outer cortical surfaces were extracted using Freesurfer and then registered into MNI space. For purposes of method development, both a classification and regression challenge were introduced for network learning including sex and age prediction, respectively. Two independent graph convolutional neural networks (gCNNs) were trained, the first of which to predict subject's self-identified sex, the second of which to predict subject's age. Class Activation Maps (CAM) and Regression Activation Maps (RAM) were constructed respectively to map the topographic distribution of the most influential brain regions involved in the decision process for each gCNN. Using this approach, the gCNN was able to predict a subject's sex with an average accuracy of 87.99 % and achieved a Person's coefficient of correlation of 0.93 with an average absolute error 4.58 years when predicting a subject's age. We believe this shape-based convolutional classifier offers a novel, data-driven approach to define biomedically relevant features from the brain at both the population and single subject levels and therefore lays a critical foundation for future precision medicine applications.

Fuzzy granular convolutional classifiers

Convolutional operations extracting effective features have been widely used in the field of deep learning. For the convolution is difficult to process set data, we propose two convolutional operators on fuzzy sets, and build a fuzzy granular classifier. Firstly, a fuzzy granulation is performed on single-atom features of classification systems to form fuzzy conditional granules and fuzzy decision granules. Then, a fuzzy conditional granular vector is constructed from the fuzzy conditional granules, and a convolutional operation is carried out on the granular vector. After that, a new fuzzy feature granule is obtained. The fuzzy feature granule is compared with its corresponding fuzzy decision granule. The result of comparison is back-propagated to the fuzzy granular vector. Simultaneously, weights of the fuzzy granular vector are modified. Thus, a fuzzy granular convolutional classifier is formed by iterating and optimizing the weights of fuzzy granular vectors several times. Furthermore, we prove the difference and derivative of fuzzy granular convolution, which provide a theoretical basis for the back-propagation of the fuzzy granular convolutional classifier. Finally, the convergence effects of the fuzzy granular convolutional operations and the classification performance of the proposed classifier are tested on some UCI datasets. The theoretical analysis and experimental results show that the convolutional operations of fuzzy granular vectors have the characteristics of fast convergence, and the fuzzy granular convolutional classifier obtains a better classification performance.

Fully automated algorithm for the detection of bone marrow oedema lesions in patients with axial spondyloarthritis – Feasibility study

In the present paper we propose an automated procedure for the detection of bone marrow oedema lesions in patients with axial spondyloarthritis (axSpA). The procedure is based on MRI examination of the sacroiliac joints of 30 patients with confirmed sacroiliitis in the course of axSpA (18 of patients were male, while 12 of patients were female; the mean age of patients was 28.8 ± 9.0 years, range from 18 to 45 years). The segmentation of the sacral and iliac bones is performed using U-Net-like architecture. The subchondral bone regions are found, where inflammatory changes are expected to appear. Convolutional classification architectures are trained to classify image voxels as either being within normal or inflammatory-changed areas. The deep learning-based classification of voxels is compared to a method based on statistical testing. The Dice coefficient for segmentation of subchondral bone was 0.84 (standard deviation 0.06). The sensitivity of the detection of inflammatory changes was 0.88. The specificity of the detection of inflammatory changes was 0.91. The discrepancy between sensitivity and specificity values achieved by the automated method and the human readers is attributed to “a satisfaction of search” phenomenon. After verification of the automated detections by human readers sensitivity and specificity increased to 0.95 and 0.96, respectively. The Spearman’s correlation coefficient between the volumes of lesions calculated manually and automatically is equal to 0.866 while the intraclass coefficient of correlation ICC(1,1) is equal to 0.947. The study demonstrates that an automated detection of inflammatory lesions with high precision of lesion volume assessment is feasible.The codes are freely available at the site:https://github.com/krzysztofrzecki/bone-marrow-oedema-detection.The data are freely available at the site:https://www.kaggle.com/krzysztofrzecki/bone-marrow-oedema-data.

Zero shot augmentation learning in internet of biometric things for health signal processing

• A novel zero shot learning method for health signal processing through data augmentation. • The method includes three parts, and a detailed workflow is designed for each part. • The method does not rely on the attribute data set. • The method explores the reasons why the classifier makes the decisions and analyzes the explanations. In recent years, the number of Internet of Things (IoT) devices has increased rapidly. The Internet of Biometric Things (IoBT) can process biometrics and health signals, and it will greatly extend the range of biometric applications. The analysis of health signals in the IoBT can use computer-aided diagnosis techniques. However, most of the existing computer-aided diagnosis methods are developed for common diseases and are not suitable for rare diseases. Zero shot learning is a potential method for the computer-aided diagnosis of rare diseases because it can identify objects of unknown categories. However, the existing zero shot learning methods are based on attribute learning and rely on an attribute dataset. There is no attribute dataset for health signal processing. Therefore, the existing zero shot learning methods are not suitable for health signal processing. Based on the above background, we propose a zero shot augmentation learning model (ZSAL) in the IoBT for health signal processing. First, an expert doctor identifies the contour of a lesion and selects a background image without a lesion. Second, the computer automatically generates virtual images using zero shot augmentation technology. Finally, the generated virtual dataset is used to train a convolutional classifier, and then we apply the classifier to the computer-aided diagnosis of actual medical images. The experiment shows the efficiency and effectiveness of our method.

An adaptive genetic algorithm as a supporting mechanism for microscopy image analysis in a cascade of convolution neural networks

The analysis of microscopy images allows for making the description of given samples containing different microscopic organisms. It is important due to the life phase analysis of these organisms. In this paper, an adaptive technique composed of a genetic algorithm (GA) and a cascade of the convolutional classifiers for image analysis is proposed. A GA is proposed for an indication of the likelihood of belonging to the appropriate class. The indicated probability is important due to the next step of measuring the results obtained from the cascade of neural classifiers. The main idea is a hybridization of these two techniques and generalization of the heuristic solution to seeking fitness function coefficients. The proposed solution is described and tested on two databases.In the case of binary classification, the obtained accuracy was 7.5% higher compared to the use of the classical approach like learning transfer. The proposed solution is discussed in relation to the analysis of individual components, other methods, and advantages/disadvantages.

Generative Enhancement of 3D Image Classifiers

In this paper, we propose a methodology for generative enhancement of existing 3D image classifiers. This methodology is based on combining the advantages of both non-generative classifiers and generative modeling. Its purpose is to streamline the synthesis of novel deep neural networks by embedding existing compatible classifiers into a generative network architecture. A demonstration of this process and evaluation of its effectiveness is performed using a 3D convolutional classifier and its generative equivalent—a 3D conditional generative adversarial network classifier. The results of the experiments show that the generative classifier delivers higher performance, gaining a relative classification accuracy improvement of 7.43%. An increase of accuracy is also observed when comparing it to a plain convolutional classifier that was trained on a dataset augmented with samples created by the trained generator. This suggests a desirable knowledge sharing mechanism exists within the hybrid discriminator-classifier network.

Residual attention-based multi-scale script identification in scene text images

Script identification is an essential step in the text extraction pipeline for multi-lingual application. This paper presents an effective approach to identify scripts in scene text images. Due to the complicated background, various text styles, character similarity of different languages, script identification has not been solved yet. Under the general classification framework of script identification, we investigate two important components: feature extraction and classification layer. In the feature extraction, we utilize a hierarchical feature fusion block to extract the multi-scale features. Furthermore, we adopt an attention mechanism to obtain the local discriminative parts of feature maps. In the classification layer, we utilize a fully convolutional classifier to generate channel-level classifications which are then processed by a global pooling layer to improve classification efficiency. We evaluated the proposed approach on benchmark datasets of RRC-MLT2017, SIW-13, CVSI-2015 and MLe2e, and the experimental results show the effectiveness of each elaborate designed component. Finally, we achieve better performances than those competitive models, where the correct rates are 89.66%, 96.11%, 98.78% and 97.20% on PRC-MLT2017, SIW-13, CVSI-2015 and MLe2e, respectively.

Deep learning method for localization and segmentation of abdominal CT

Computed Tomography (CT) imaging is widely used for studying body composition, i.e., the proportion of muscle and fat tissues with applications in areas such as nutrition or chemotherapy dose design. In particular, axial CT slices from the 3rd lumbar (L3) vertebral location are commonly used for body composition analysis. However, selection of the third lumbar vertebral slice and the segmentation of muscle/fat in the slice is a tedious operation if performed manually. The objective of this study is to automatically find the middle axial slice at L3 level from a full or partial body CT scan volume and segment the skeletal muscle (SM), subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT) and intermuscular adipose tissue (IMAT) on that slice. The proposed algorithm includes an L3 axial slice localization network followed by a muscle-fat segmentation network. The localization network is a fully convolutional classifier trained on more than 12,000 images. The segmentation network is a convolutional neural network with an encoder–decoder architecture. Three datasets with CT images taken for patients with different types of cancers are used for training and validation of the networks. The mean slice error of 0.87±2.54 was achieved for L3 slice localization on 1748 CT scan volumes. The performance of five class tissue segmentation network evaluated on two datasets with 1327 and 1202 test samples. The mean Jaccard score of 97% was achieved for SM and VAT tissue segmentation on 1327 images. The mean Jaccard scores of 98% and 83% are corresponding to SAT and IMAT tissue segmentation on the same dataset. The localization and segmentation network performance indicates the potential for fully automated body composition analysis with high accuracy.

Deep Neural Framework with Visual Attention and Global Context for Predicting Image Aesthetics

Computational inference of aesthetics has recently become a hot topic due to its usefulness in widely applications such as evaluating image quality, retouching image and retrieving image. Owing to the subjectivity of this problem, there is no general framework to predict image aesthetics. In this paper, we propose a deep neural framework with visual attention module, self-generated global features and hybrid loss to address this problem. Specifically, the framework can be any state-of-the-art convolution classification network compatible with visual attention. Further, self-generated global feature compensates for the loss of global context information during training stage, and the hybrid loss guides the network to learn the similarity between the predicted aesthetic scores and the ground-truths through fusing soft-max-entropy and Earth Mover’s Distance(EMD). With the above-mentioned improvements, the proposed deep neural framework is capable of effectively predicting image aesthetics in an efficient way. In our experiments, we release a real-world aesthetic dataset that contains 1,800 2K photos labeled by several experienced photographers, and then provide a thorough ablation study of the design choices to better understand the superiority brought by each part of our framework, and design several comparisons with the state-of-the-art methods on a fraction of metrics. The experimental results on two datasets demonstrate that both accuracy and efficiency achieve favorably performance.

Context propagation embedding network for weakly supervised semantic segmentation

Weakly-supervised semantic segmentation with image-level labels has much significance as it facilitates related practical applications under lightweight manual annotations. Recent researches first infer the visual cues, referred as discriminative regions corresponding to each object category on images using deep convolutional classification networks. Then they expand visual cues to generate initial segmentation masks. Despite the remarkable progress, the segmentation performance still remains unsatisfactory due to the absence of complete visual cues. Using these low-quality visual cues as prior will have the limitation on improving segmentation performance. To overcome this problem, we propose a novel context propagation embedding network, i.e., the CPENet to generate high-quality visual cues, which focuses on learning semantic relationship between each region and its surrounding neighbors and selectively propagates discriminative information to non-discriminative, object related regions. Our methods can provide reliable initial segmentation masks for training subsequent segmentation network to generate final segmentation results. In addition, we refined convolutional block attention module (CBAM) [30] to hierarchically extract more category-aware features by capturing global context related information and further promote the propagation process. Experiments on benchmark demonstrate that our proposed method obtains superior performance over the state-of-the-arts.

An Intellectual Detection System for Intrusions based on Collaborative Machine Learning

The necessity for safety of information in a network has inflated due to the impressive growth of web applications. Several methods of intrusion detection are used to detect irregularities which depend on precision, detection frequency, other parameters and are anticipated to familiarize to vigorously varying risk scenes. To accomplish consistent abnormalities detection in a network many machine learning algorithms have been formulated by researchers. A technique based on unsupervised machine learning that use two separate machine learning algorithms to identify anomalies in a network viz convolutional autoencoder and softmax classifier is proposed. These profound models were skilled as well as evaluated on NSLKDD test data sets on the NSLKDD training dataset. Using well-known classification metrics such as accuracy, precision and recall, these machine learning models were assessed. The developed intrusion detection system model experimental findings showed promising outcomes in anomaly detection systems for real-world implementation and is compared with the prevailing definitive machine learning techniques. This strategy increases the detection of network intrusion and offers a renewed intrusion detection study method.

Granule Vectors and Granular Convolutional Classifiers

Convolutional operations can extract effective features and have been widely used in the field of deep learning. For the deficiency of convolution mainly dealing with numerical data, we propose a novel convolutional operator on granules with a set form, further we build a classifier on it. Firstly, feature granules are constructed on each single feature of a classification system by introducing neighborhood rough sets. Synchronously, decision granules are generated on the labels of samples. Secondly, feature granule vectors and weighted granule vectors are constructed from these granules, and a convolutional operation is proposed on feature granule vectors and weighted granule vectors, then a predicted granule is produced as a result of the convolutional operation. The predicted granule is compared with the decision granule, and their residual error is back propagated to the weighted granule vector for tuning its value. After multiple iterations of the granular convolutional operations and back propagation corrections, the weight of the granular vector is convergent and optimized. Furthermore, a granular classifier is designed based on the convolutional operation. The constringency of the granular convolution and the classification performance of the granular classifier are tested on some UCI datasets. Theoretical analysis and experimental results show that the granular convolution has a characteristic of fast convergence, and the granular convolutional classifier has a better classification performance.

Segmentation of shoulder rehabilitation exercises for single and multiple inertial sensor systems.

IntroductionDigital home rehabilitation systems require accurate segmentation methods to provide appropriate feedback on repetition counting and exercise technique. Current segmentation methods are not suitable for clinical use; they are not highly accurate or require multiple sensors, which creates usability problems. We propose a model for accurately segmenting inertial measurement unit data for shoulder rehabilitation exercises. This study aims to use inertial measurement unit data to train and test a machine learning segmentation model for single- and multiple-inertial measurement unit systems and to identify the optimal single-sensor location.MethodsA focus group of specialist physiotherapists selected the exercises, which were performed by participants wearing inertial measurement units on the wrist, arm and scapula. We applied a novel machine learning based segmentation technique involving a convolutional classifier and Finite State Machine to the inertial measurement unit data. An accuracy score was calculated for each possible single- or multiple-sensor system.ResultsThe wrist inertial measurement unit was chosen as the optimal single-sensor location for future system development (mean overall accuracy 0.871). Flexion and abduction based exercises mostly could be segmented with high accuracy, but scapular movement exercises had poor accuracy.ConclusionA wrist-worn single inertial measurement unit system can accurately segment shoulder exercise repetitions; however, accuracy varies depending on characteristics of the exercise.

Imbalanced Fault Diagnosis of Rolling Bearing Using Enhanced Generative Adversarial Networks

Machinery fault diagnosis tasks have been well addressed when sufficient and abundant data are available. However, the data imbalance problem widely exists in real-world scenarios, which leads to the performance deterioration of fault diagnosis markedly. To solve this problem, we present a novel imbalanced fault diagnosis method based on the enhanced generative adversarial networks (GAN). By artificially generating fake samples, the proposed method can mitigate the loss caused by the lack of real fault data. Specifically, in order to improve the quality of generated samples, a new discriminator is designed using spectrum normalization (SN) strategy and a two time-scale update rule (TTUR) method is used to stabilize the training process of GAN. Then, an enhanced Wasserstein GAN with gradient penalty is developed to generate high-quality synthetic samples for the fault samples set. Finally, a deep convolutional classifier is constructed to carry out fault classification. The performance and effectiveness of the proposed method are validated on the Case Western Reserve University bearing dataset and rolling bearing dataset acquired from our laboratory. The simulation results show that the proposed method has a superior performance than other methods for imbalanced fault diagnosis tasks.

Bearing Defect Classification Algorithm Based on Autoencoder Neural Network

The postproduction defect classification and detection of bearings still relies on manual detection, which is time‐consuming and tedious. To address this, we propose a bearing defect classification network based on an autoencoder to enhance the efficiency and accuracy of bearing defect detection. An improved autoencoder is used to reduce dimension feature extraction and reduce large‐scale images to small‐scale images through encoder dimensional reduction. Defect classification is completed by feeding the extracted features into a convolutional classification network. Comparative experiments show that the neural network can effectively complete feature selection and substantially improve classification accuracy while avoiding the laborious algorithm of the conventional method.