Shallow Convolutional Neural Network Research Articles

Uncertainty measurement-guided iterative sample selection via shallow convolutional neural network for hyperspectral image classification

Continuously improving the accuracy is a hot topic in hyperspectral image (HSI) classification with small-scale samples, due to the high label noise of traditional labeling systems and the high cost of expert labeling systems. We focus on constructing a smaller and more informative training sample set, so an iterative sample selection method guided by uncertainty measurement (ISS-Un) is proposed. The method learns shallow and deep features in the spectral and spatial domains via a convolutional neural network (CNN), where an uncertainty measurement algorithm such as least confidence (LC), marginal sampling (MS) or entropy (Ent) is used to iteratively select high-quality samples for the training set. In addition, we propose a more efficient uncertainty measurement algorithm named margin-entropy fusion (MEF) algorithm to integrate multiple-criteria information. The proposed method is compared with the conventional random sampling method. Experimental results on three HSI datasets show that the proposed ISS-Un method can significantly alleviate the redundancy of training samples and form a more compact and efficient training set, thus improving the classification performance of pixel-oriented HSI. Meanwhile, training sets constructed based on different uncertainty measurement algorithms are applied to five popular CNN models to verify the quality and generalizability of the selected samples. The results show that these training sets work better than random training sampling. Moreover, the proposed MEF algorithm outperforms the LC, MS, and Ent algorithms in selecting samples and is the main recommended scheme.

Mental arithmetic task classification with convolutional neural network based on spectral-temporal features from EEG.

In recent years, neuroscientists have been interested to the development of brain-computer interface (BCI) devices. Patients with motor disorders may benefit from BCIs as a means of communication and for the restoration of motor functions. Electroencephalography (EEG) is one of most used for evaluating the neuronal activity. In many computer vision applications, deep neural networks (DNN) show significant advantages. Towards to ultimate usage of DNN, we present here a shallow neural network that uses mainly two convolutional neural network (CNN) layers, with relatively few parameters and fast to learn spectral-temporal features from EEG. We compared this models to three other neural network models with different depths applied to a mental arithmetic task using eye-closed state adapted for patients suffering from motor disorders and a decline in visual functions. Experimental results showed that the shallow CNN model outperformed all the other models and achieved the highest classification accuracy of 90.68%. It's also more robust to deal with cross-subject classification issues: only 3% standard deviation of accuracy instead of 15.6% from conventional method.

Modified SSR-NET: A Shallow Convolutional Neural Network for Efficient Hyperspectral Image Super-Resolution

A fast and shallow convolutional neural network is proposed for hyperspectral image super-resolution inspired by Spatial-Spectral Reconstruction Network (SSR-NET). The feature extraction ability is improved compared to SSR-NET and other state-of-the-art methods, while the proposed network is also shallow. Numerical experiments show both the visual and quantitative superiority of our method. Specifically, for the fusion setup with two inputs, obtained by 32× spatial downsampling for the low-resolution hyperspectral (LR HSI) input and 25× spectral downsampling for high-resolution multispectral (HR MSI) input, a significant improvement of the quality of super-resolved HR HSI over 4 dB is demonstrated as compared with SSR-NET. It is also shown that, in some cases, our method with a single input, HR MSI, can provide a comparable result with that achieved with two inputs, HR MSI and LR HSI.

Performance evaluation of shallow and deep CNN architectures on building segmentation from high-resolution images

Performance evaluation of shallow and deep CNN architectures on building segmentation from high-resolution images

Facial Expressions Based Automatic Pain Assessment System

Pain assessment is used to improve patients’ treatment outcomes. Human observers may be influenced by personal factors, such as inexperience and medical organizations are facing a shortage of experts. In this study, we developed a facial expressions-based automatic pain assessment system (FEAPAS) to notify medical staff when a patient suffers pain by activating an alarm and recording the incident and pain level with the date and time. The model consists of two identical concurrent subsystems, each of which takes one of the two inputs of the model, i.e., “full face” and “the upper half of the same face”. The subsystems extract the relevant input features via two pre-trained convolutional neural networks (CNNs), using either VGG16, InceptionV3, ResNet50, or ResNeXt50, while freezing all convolutional blocks and replacing the classifier layer with a shallow CNN. The concatenated outputs in this stage is then sent to the model’s classifier. This approach mimics the human observer method and gives more importance to the upper part of the face, which is similar to the Prkachin and Soloman pain intensity (PSPI). Additionally, we further optimized our models by applying four optimizers (SGD/ADAM/RMSprop/RAdam) to each model and testing them on the UNBC-McMaster shoulder pain expression archive dataset to find the optimal combination, InceptionV3-SGD. The optimal model showed an accuracy of 99.10% on 10-fold cross-validation, thus outperforming the state-of-the-art model on the UNBC-McMaster database. It also scored 90.56% on unseen subject data. To speed up the system response time and reduce unnecessary alarms associated with temporary facial expressions, a select but effective subset of frames was inspected and classified. Two frame-selection criteria were reported. Classifying only two frames at the middle of 30-frame sequence was optimal, with an average reaction time of at most 6.49 s and the ability to avoid unnecessary alarms.

Reference-free learning-based similarity metric for motion compensation in cone-beam CT

Purpose. Patient motion artifacts present a prevalent challenge to image quality in interventional cone-beam CT (CBCT). We propose a novel reference-free similarity metric (DL-VIF) that leverages the capability of deep convolutional neural networks (CNN) to learn features associated with motion artifacts within realistic anatomical features. DL-VIF aims to address shortcomings of conventional metrics of motion-induced image quality degradation that favor characteristics associated with motion-free images, such as sharpness or piecewise constancy, but lack any awareness of the underlying anatomy, potentially promoting images depicting unrealistic image content. DL-VIF was integrated in an autofocus motion compensation framework to test its performance for motion estimation in interventional CBCT. Methods. DL-VIF is a reference-free surrogate for the previously reported visual image fidelity (VIF) metric, computed against a motion-free reference, generated using a CNN trained using simulated motion-corrupted and motion-free CBCT data. Relatively shallow (2-ResBlock) and deep (3-Resblock) CNN architectures were trained and tested to assess sensitivity to motion artifacts and generalizability to unseen anatomy and motion patterns. DL-VIF was integrated into an autofocus framework for rigid motion compensation in head/brain CBCT and assessed in simulation and cadaver studies in comparison to a conventional gradient entropy metric. Results. The 2-ResBlock architecture better reflected motion severity and extrapolated to unseen data, whereas 3-ResBlock was found more susceptible to overfitting, limiting its generalizability to unseen scenarios. DL-VIF outperformed gradient entropy in simulation studies yielding average multi-resolution structural similarity index (SSIM) improvement over uncompensated image of 0.068 and 0.034, respectively, referenced to motion-free images. DL-VIF was also more robust in motion compensation, evidenced by reduced variance in SSIM for various motion patterns (σ DL-VIF = 0.008 versus σ gradient entropy = 0.019). Similarly, in cadaver studies, DL-VIF demonstrated superior motion compensation compared to gradient entropy (an average SSIM improvement of 0.043 (5%) versus little improvement and even degradation in SSIM, respectively) and visually improved image quality even in severely motion-corrupted images.Conclusion: The studies demonstrated the feasibility of building reference-free similarity metrics for quantification of motion-induced image quality degradation and distortion of anatomical structures in CBCT. DL-VIF provides a reliable surrogate for motion severity, penalizes unrealistic distortions, and presents a valuable new objective function for autofocus motion compensation in CBCT.

Accurate modulation classification under impaired wireless channels via shallow convolutional neural networks

Classifying the modulation type of radio signals plays an important role in current and future wireless communication systems. We present a modulation classification method based on convolutional neural networks that reaches high accuracy in face of various channel characteristics and signal conditions without requiring the network to have a very large depth. Experiment results show that the proposed method reaches accurate classification under different system impairment settings that include sampling rate offset, carrier frequency offset, multi-path fading, and additive white Gaussian noise. For instance, compared to a state-of-the-art method, accuracy is improved up to 25% in classifying difficult modulation types under system impairments. Source code of the proposed method is available online.

An effective hashing method using W-Shaped contrastive loss for imbalanced datasets

The extraction of informative features from medical images and the retrieving of similar images from data repositories is vital for clinical decision support systems. Unlike general tasks such as medical image classification and segmentation, retrieval is more reliable in terms of interpretability. However, this task is quite challenging due to the multimodal and imbalanced nature of medical images. Because traditional retrieval methods use hand-crafted feature extraction guided approximate hashing functions, they often have problems capturing the latent characteristics of images. Deep learning based retrieval methods can eliminate drawbacks of hand-crafted feature extraction methods. However, in order for a deep architecture to produce high performance, large-scale datasets containing labeled and balanced samples are required. Since most medical datasets do not have these properties, existing hashing methods are not powerful enough to model patterns in medical images, which have a similar general appearance but subtle differences. In this study, a novel W-shaped contrastive loss (W-SCL) is proposed for skin lesion image retrieval on a dataset whose visual difference between classes is relatively low. We considerably improve the traditional contrastive loss (CL) performance by including label information for very similar skin lesion images. We use two benchmark datasets consisting of general images and two benchmark skin lesion datasets to test the proposed W-SCL performance. In addition, experiments are carried out using various pre-trained CNN and shallow CNN architectures. These extensive experiments reveal that the proposed method improves the mean average precision (mAP) performance by approximately 7% for general image datasets and approximately 12% for skin lesion datasets.

Super-Pixel Guided Low-Light Images Enhancement with Features Restoration.

Dealing with low-light images is a challenging problem in the image processing field. A mature low-light enhancement technology will not only be conductive to human visual perception but also lay a solid foundation for the subsequent high-level tasks, such as target detection and image classification. In order to balance the visual effect of the image and the contribution of the subsequent task, this paper proposes utilizing shallow Convolutional Neural Networks (CNNs) as the priori image processing to restore the necessary image feature information, which is followed by super-pixel image segmentation to obtain image regions with similar colors and brightness and, finally, the Attentive Neural Processes (ANPs) network to find its local enhancement function on each super-pixel to further restore features and details. Through extensive experiments on the synthesized low-light image and the real low-light image, the experimental results of our algorithm reach 23.402, 0.920, and 2.2490 for Peak Signal to Noise Ratio (PSNR), Structural Similarity (SSIM), and Natural Image Quality Evaluator (NIQE), respectively. As demonstrated by the experiments on image Scale-Invariant Feature Transform (SIFT) feature detection and subsequent target detection, the results of our approach achieve excellent results in visual effect and image features.

Rolling Bearing Fault Diagnosis in Limited Data Scenarios Using Feature Enhanced Generative Adversarial Networks

Most current rolling bearing fault diagnostic approaches presume that sufficient labeled data are available for training. However, labeled fault data collection in realistic engineering is time-consuming and expensive, whereas the majority of data available are unlabeled, restricting diagnostic performance. To alleviate the dependence of feature self-extraction model on rolling bearing fault data and improve the feature capture ability, a bearing fault diagnosis method based on feature-enhanced generative adversarial networks with an auxiliary classifier (AC-FEGAN) is proposed in this paper. Firstly, the non-saturating loss with gradient penalty term is introduced to circumvent the gradient vanishing and improve learning stability. Then, employing identity mapping of residual networks introduced into AC-FEGAN, representative faulty features can be automatically extracted, overcoming the restrictions of shallow convolutional neural networks in feature extraction. Simultaneously, to improve feature learning ability, the self-attention module embedded in the residual network accomplishes refined feature extraction by incorporating appropriate weight matrices into the faulty feature maps. Finally, the auxiliary classifier is employed as a pre-training model for the fault diagnosis model, with generated and unlabeled samples employed to fine-tune the auxiliary classifier to identify faulty bearings. The experimental results show that the proposed method can significantly improve the quality of generated samples while eliminating the reliance of existing diagnosis methods on data volume to achieve more accurate and effective fault diagnosis.

Pedestrian and Animal Recognition Using Doppler Radar Signature and Deep Learning.

Pedestrian occurrences in images and videos must be accurately recognized in a number of applications that may improve the quality of human life. Radar can be used to identify pedestrians. When distinct portions of an object move in front of a radar, micro-Doppler signals are produced that may be utilized to identify the object. Using a deep-learning network and time–frequency analysis, we offer a method for classifying pedestrians and animals based on their micro-Doppler radar signature features. Based on these signatures, we employed a convolutional neural network (CNN) to recognize pedestrians and animals. The proposed approach was evaluated on the MAFAT Radar Challenge dataset. Encouraging results were obtained, with an AUC (Area Under Curve) value of 0.95 on the public test set and over 0.85 on the final (private) test set. The proposed DNN architecture, in contrast to more common shallow CNN architectures, is one of the first attempts to use such an approach in the domain of radar data. The use of the synthetic radar data, which greatly improved the final result, is the other novel aspect of our work.

ECG-iCOVIDNet: Interpretable AI model to identify changes in the ECG signals of post-COVID subjects

ObjectiveStudies showed that many COVID-19 survivors develop sub-clinical to clinical heart damage, even if subjects did not have underlying heart disease before COVID. Since Electrocardiogram (ECG) is a reliable technique for cardiovascular disease diagnosis, this study analyzes the 12-lead ECG recordings of healthy and post-COVID (COVID-recovered) subjects to ascertain ECG changes after suffering from COVID-19. MethodWe propose a shallow 1-D convolutional neural network (CNN) deep learning architecture, namely ECG-iCOVIDNet, to distinguish ECG data of post-COVID subjects and healthy subjects. Further, we employed ShAP technique to interpret ECG segments that are highlighted by the CNN model for the classification of ECG recordings into healthy and post-COVID subjects. ResultsECG data of 427 healthy and 105 post-COVID subjects were analyzed. Results show that the proposed ECG-iCOVIDNet model could classify the ECG recordings of healthy and post-COVID subjects better than the state-of-the-art deep learning models. The proposed model yields an F1-score of 100%. ConclusionSo far, we have not come across any other study with an in-depth ECG signal analysis of the COVID-recovered subjects. In this study, it is shown that the shallow ECG-iCOVIDNet CNN model performed good for distinguishing ECG signals of COVID-recovered subjects from those of healthy subjects. In line with the literature, this study confirms changes in the ECG signals of COVID-recovered patients that could be captured by the proposed CNN model. Successful deployment of such systems can help the doctors identify the changes in the ECG of the post-COVID subjects on time that can save many lives.

Retinopathy grading with deep learning and wavelet hyper-analytic activations.

Recent developments reveal the prominence of Diabetic Retinopathy (DR) grading. In the past few decades, Wavelet-based DR classification has shown successful impacts and the Deep Learning models, like Convolutional Neural Networks (CNN’s), have evolved in offering the highest prediction accuracy. In this work, the features of the input image are enhanced with the integration of Multi-Resolution Analysis (MRA) and a CNN framework without costing more convolution filters. The bottleneck with conventional activation functions, used in CNN’s, is the nullification of the feature maps that are negative in value. In this work, a novel Hyper-analytic Wavelet (HW) phase activation function is formulated with unique characteristics for the wavelet sub-bands. Instead of dismissal, the function transforms these negative coefficients that correspond to significant edge feature maps. The hyper-analytic wavelet phase forms the imaginary part of the complex activation. And the hyper-parameter of the activation function is selected such that the corresponding magnitude spectrum produces monotonic and effective activations. The performance of 3 CNN models (1 custom, shallow CNN, ResNet with Soft attention, Alex Net for DR) with spatial–Wavelet quilts is better. With the spatial–Wavelet quilts, the Alex Net for DR has an improvement with an 11% of accuracy level (from 87 to 98%). The highest accuracy level of 98% and the highest Sensitivity of 99% are attained through Modified Alex Net for DR. The proposal also illustrates the visualization of the negative edge preservation with assumed image patches. From this study, the researcher infers that models with spatial–Wavelet quilts, with the hyper-analytic activations, have better generalization ability. And the visualization of heat maps provides evidence of better learning of the feature maps from the wavelet sub-bands.

Recognition of void defects in airport runways using ground-penetrating radar and shallow CNN

The interpretation of ground-penetrating radar (GPR) data usually depends on manual work, thus yielding unsatisfactory accuracy when processing complicated cases. To address this issue, the finite-difference time-domain method was adopted to explore the void-related features in GPR images, and guide the field survey to locate the void area. On this basis, a dataset including 811 void features was established from 10 airport runways using an 800 MHz ground-coupled GPR system. Data augmentation was used to enlarge the dataset and train four configurations of a mixed convolution neural network (CNN) model with feature extractors (ResNet18 and ResNet50) and object detectors (YOLOv2 and Faster R-CNN). To make the detection process automatic, an incremental random sampling (IRS) approach was used to generate images from GPR data and then fed into the trained model. The experimental results demonstrated that the shallow ResNet18-YOLOv2 with the IRS method is a promising strategy for void detection in airport runways.

Decoding continuous kinetic information of grasp from stereo-electroencephalographic (SEEG) recordings

Objective. Brain–computer interfaces (BCIs) have the potential to bypass damaged neural pathways and restore functionality lost due to injury or disease. Approaches to decoding kinematic information are well documented; however, the decoding of kinetic information has received less attention. Additionally, the possibility of using stereo-electroencephalography (SEEG) for kinetic decoding during hand grasping tasks is still largely unknown. Thus, the objective of this paper is to demonstrate kinetic parameter decoding using SEEG in patients performing a grasping task with two different force levels under two different ascending rates. Approach. Temporal-spectral representations were studied to investigate frequency modulation under different force tasks. Then, force amplitude was decoded from SEEG recordings using multiple decoders, including a linear model, a partial least squares model, an unscented Kalman filter, and three deep learning models (shallow convolutional neural network, deep convolutional neural network and the proposed CNN+RNN neural network). Main results. The current study showed that: (a) for some channel, both low-frequency modulation (event-related desynchronization (ERD)) and high-frequency modulation (event-related synchronization) were sustained during prolonged force holding periods; (b) continuously changing grasp force can be decoded from the SEEG signals; (c) the novel CNN+RNN deep learning model achieved the best decoding performance, with the predicted force magnitude closely aligned to the ground truth under different force amplitudes and changing rates. Significance. This work verified the possibility of decoding continuously changing grasp force using SEEG recordings. The result presented in this study demonstrated the potential of SEEG recordings for future BCI application.

A deep learning-based framework for detecting COVID-19 patients using chest X-rays.

Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) has caused outbreaks of new coronavirus disease (COVID-19) around the world. Rapid and accurate detection of COVID-19 coronavirus is an important step in limiting the spread of the COVID-19 epidemic. To solve this problem, radiography techniques (such as chest X-rays and computed tomography (CT)) can play an important role in the early prediction of COVID-19 patients, which will help to treat patients in a timely manner. We aimed to quickly develop a highly efficient lightweight CNN architecture for detecting COVID-19-infected patients. The purpose of this paper is to propose a robust deep learning-based system for reliably detecting COVID-19 from chest X-ray images. First, we evaluate the performance of various pre-trained deep learning models (InceptionV3, Xception, MobileNetV2, NasNet and DenseNet201) recently proposed for medical image classification. Second, a lightweight shallow convolutional neural network (CNN) architecture is proposed for classifying X-ray images of a patient with a low false-negative rate. The data set used in this work contains 2,541 chest X-rays from two different public databases, which have confirmed COVID-19 positive and healthy cases. The performance of the proposed model is compared with the performance of pre-trained deep learning models. The results show that the proposed shallow CNN provides a maximum accuracy of 99.68% and more importantly sensitivity, specificity and AUC of 99.66%, 99.70% and 99.98%. The proposed model has fewer parameters and low complexity compared to other deep learning models. The experimental results of our proposed method show that it is superior to the existing state-of-the-art methods. We believe that this model can help healthcare professionals to treat COVID-19 patients through improved and faster patient screening.

ODGNet: a deep learning model for automated optic disc localization and glaucoma classification using fundus images

Glaucoma is one of the prevalent causes of blindness in the modern world. It is a salient chronic eye disease that leads to irreversible vision loss. The impediments of glaucoma can be restricted if it is identified at primary stages. In this paper, a novel two-phase Optic Disk localization and Glaucoma Diagnosis Network (ODGNet) has been proposed. In the first phase, a visual saliency map incorporated with shallow CNN is used for effective OD localization from the fundus images. In the second phase, the transfer learning-based pre-trained models are used for glaucoma diagnosis. The transfer learning-based models such as AlexNet, ResNet, and VGGNet incorporated with saliency maps are evaluated on five public retinal datasets (ORIGA, HRF, DRIONS-DB, DR-HAGIS, and RIM-ONE) to differentiate between normal and glaucomatous images. This study’s experimental results demonstrate that the proposed ODGNet evaluated on ORIGA for glaucoma diagnosis is the most predictive model and achieve 95.75, 94.90, 94.75, and 97.85% of accuracy, specificity, sensitivity, and area under the curve, respectively. These results indicate that the proposed OD localization method based on the saliency map and shallow CNN is robust, accurate and saves the computational cost.

Identification and quantification of adulterated honey by Raman spectroscopy combined with convolutional neural network and chemometrics

In this study, Raman spectroscopy combined with convolutional neural network (CNN) and chemometrics was used to achieve the identification and quantification of honey samples adulterated with high fructose corn syrup, rice syrup, maltose syrup and blended syrup, respectively. The shallow CNNs utilized to analyze honey mixed with single-variety syrup classified samples into four categories by the adulteration concentration with more than 97% accuracy, and the general CNN model for simultaneously detecting honey adulterated with any type of syrup obtained an accuracy of 94.79%. The established CNNs had the best performance compared with several chemometric classification algorithms. In addition, partial least square regression (PLS) successfully predicted the purity of honey mixed with single syrup, while coefficients of determination and root mean square errors of prediction were greater than 0.98 and less than 3.50, respectively. Therefore, the proposed methods based on Raman spectra have important practical significance for food safety and quality control of honey products.

Efficient optimisation framework for convolutional neural networks with secure multiparty computation

In recent years, deep learning has become an increasingly popular approach to modelling data due to its ability to detect abstract underlying patterns in data. Its practical applications have been limited, however, by data privacy concerns, restricting its use in major sectors such as healthcare and banking. Secure multiparty computation (MPC) is a scheme which allows multiple parties to perform joint computations over private data, while keeping the content of their data secret. MPC can enable privacy-preserving machine learning, however current implementations are rarely applied in practice due to the prohibitively high cost of performing thousands of computations and transmitting data between parties. In this paper we propose a framework incorporating various optimisation approaches from the wider field of privacy-preserving deep learning, including privacy-preserving batch normalisation and polynomial approximation of activation functions, and evaluate their performance when applied to a privacy-preserving convolutional neural network (CNN), discussing the trade-off each offers in terms of their accuracy and efficiency. We experiment with parametric polynomial (PPoly) activations by deriving polynomial approximations to activation functions and allowing the network to tune the coefficients as learning weights. We will show that, in shallow CNNs, the application of batch normalisation in combination with a PPoly activation layer can result in faster convergence, with testing accuracy exceeding that achieved with an unencrypted network, at the cost of longer running times.

Pan-Sharpening Based on CNN+ Pyramid Transformer by Using No-Reference Loss

The majority of existing deep learning pan-sharpening methods often use simulated degraded reference data due to the missing of real fusion labels which affects the fusion performance. The normally used convolutional neural network (CNN) can only extract the local detail information well which may cause the loss of important global contextual characteristics with long-range dependencies in fusion. To address these issues and to fuse spatial and spectral information with high quality information from the original panchromatic (PAN) and multispectral (MS) images, this paper presents a novel pan-sharpening method by designing the CNN+ pyramid Transformer network with no-reference loss (CPT-noRef). Specifically, the Transformer is used as the main architecture for fusion to supply the global features, the local features in shallow CNN are combined, and the multi-scale features from the pyramid structure adding to the Transformer encoder are learned simultaneously. Our loss function directly learns the spatial information extracted from the PAN image and the spectral information from the MS image which is suitable for the theory of pan-sharpening and makes the network control the spatial and spectral loss simultaneously. Both training and test processes are based on real data, so the simulated degraded reference data is no longer needed, which is quite different from most existing deep learning fusion methods. The proposed CPT-noRef network can effectively solve the huge amount of data required by the Transformer network and extract abundant image features for fusion. In order to assess the effectiveness and universality of the fusion model, we have trained and evaluated the model on the experimental data of WorldView-2(WV-2) and Gaofen-1(GF-1) and compared it with other typical deep learning pan-sharpening methods from both the subjective visual effect and the objective index evaluation. The results show that the proposed CPT-noRef network offers superior performance in both qualitative and quantitative evaluations compared with existing state-of-the-art methods. In addition, our method has the strongest generalization capability by testing the Pleiades and WV-2 images on the network trained by GF-1 data. The no-reference loss function proposed in this paper can greatly enhance the spatial and spectral information of the fusion image with good performance and robustness.