Popular Convolutional Neural Networks Research Articles

FDNet: Imperceptible backdoor attacks via frequency domain steganography and negative sampling

Backdoor attacks against Deep Neural Networks (DNNs) have surfaced as a substantial and concerning security challenge. These backdoor vulnerabilities in DNNs can be introduced by third-party sources through maliciously manipulated training data. Existing backdoor attacks are primarily built on perturbation trigger patterns in the spatial domain, which makes practical deployment arduous due to the ease of detection by inspectors. Moreover, shortcut learning renders the backdoor network less robust against defense methods. This work advances an effective and adaptable approach to backdoor attacks, situated in the frequency domain. This methodology involves the incorporation of specific natural perturbations within the frequency domain of images. Remarkably, these introduced triggers yield minimal alterations in the image’s semantic content, rendering them nearly imperceptible to human observers. To evade detection by machine-based defenders, we introduce a new training paradigm that incorporates negative sampling techniques. This approach compels the neural network to learn richer differences as trigger patterns. We evaluate our attacks on popular convolutional neural networks, visual transformers, and MLP-Mixer models as well as four standard datasets including MNIST, CIFAR-10, GTSRB, and ImageNet. Experimental results demonstrate that the trained networks can be successfully injected with backdoors. Our attack methods exhibit remarkable efficacy, achieving high attack success rates in both All-to-one (near 100% on all datasets) and All-to-all (over 90% except on ImageNet) scenarios and also demonstrate robustness against contemporary state-of-the-art defense mechanisms. Furthermore, our work reveals that DNNs can capture discrepancies in the frequency components of images that are barely perceptible to humans.

Automatic stack velocity picking using a semi‐supervised ensemble learning method

AbstractPicking stack velocity from seismic velocity spectra is a fundamental method in seismic stack velocity analysis. With the increase in the scale of seismic data acquisition, manual picking cannot achieve the required efficiency. Therefore, an automatic picking algorithm is urgently needed now. Despite some supervised deep learning–based picking approaches that have been proposed, they heavily rely on sufficient training samples and lack interpretability. In contrast, utilizing physical knowledge to develop semi‐data‐driven methods has the potential to efficiently solve this problem. Thus, we propose a semi‐supervised ensemble learning method to reduce the reliance on manually labelled data and improve interpretability by incorporating the interval velocity constraint. Semi‐supervised ensemble learning fuses the information of the estimated spectrum, nearby velocity spectra and few‐shot manual picking to recognize the velocity picking. Test results of both the synthetic and field datasets indicate that semi‐supervised ensemble learning achieves more reliable and precise picking than traditional clustering‐based techniques and the currently popular convolutional neural network method.

A CNN-based model to count the leaves of rosette plants (LC-Net)

Plant image analysis is a significant tool for plant phenotyping. Image analysis has been used to assess plant trails, forecast plant growth, and offer geographical information about images. The area segmentation and counting of the leaf is a major component of plant phenotyping, which can be used to measure the growth of the plant. Therefore, this paper developed a convolutional neural network-based leaf counting model called LC-Net. The original plant image and segmented leaf parts are fed as input because the segmented leaf part provides additional information to the proposed LC-Net. The well-known SegNet model has been utilised to obtain segmented leaf parts because it outperforms four other popular Convolutional Neural Network (CNN) models, namely DeepLab V3+, Fast FCN with Pyramid Scene Parsing (PSP), U-Net, and Refine Net. The proposed LC-Net is compared to the other recent CNN-based leaf counting models over the combined Computer Vision Problems in Plant Phenotyping (CVPPP) and KOMATSUNA datasets. The subjective and numerical evaluations of the experimental results demonstrate the superiority of the LC-Net to other tested models.

Enhanced symbiotic organism search optimization algorithm for plant disease classification

Since it satisfies all prerequisites for the growth of humanity, agriculture is currently regarded as being the most significant sector for civilization. One of the main forms of human energy production is thought to be plants, which also provide nutrients, cures, etc. Any damage or disease brought on by exposure to pathogens, viruses, bacteria, etc., while cultivating plants results in a decline in productivity, making it crucial to prevent such diseases and take the required precautions to avoid them. Accurately identifying such fatal diseases is a crucial first step for both the businesses and farmers. Six different Convolutional Neural Networks (CNNs) that accept plant leaf images as input, along with the Enhanced Symbiotic Organism Search (ESOS) optimization algorithm, have been implemented in our research. We intend to extensively contrast the various models based on accuracy, precision, recall, and F1-score. In the area of image recognition and classification, convolutional neural networks (CNNs), in particular, and deep learning, in general, are developing. The literature contains a variety of CNN designs. The dataset size, the number of classes, the model’s weights, hypermeters, and optimizers are a few examples of the variables that have an impact on a CNN model’s performance. Because of its benefits, transfer learning and fine-tuning a pre-trained model are now very popular. This study examines the impact of six popular CNN models: DenseNet, MobileNet, EfficientNet, VGG19, ResNet and Inception. As a result, DenseNet demonstrates an optimal accuracy rate of 98% when compared to other models.

From Detection to Action: Implementing Deep Learning Inference in PLC Systems via Docker Services

This paper focuses on the implementation of recognition and detection algorithms for programmable logic controllers (PLCs) based on artificial intelligence (AI). AI methods based on deep neural networks (DNN) are characterized by a large number of hidden layers and neurons in hidden layers, which increases the computational complexity of computing the outputs of artificial neural networks (ANNs). Popular convolutional neural networks (CNN) are among such ResNet, AlexNet, GoogLeNet and others. A remote cloud solution is proposed in this paper to compute CNN outputs. The CNN implementation was tested on object recognition and detection applications. Training and testing of the CNN and YOLO detector was carried out in the Python environment.

Image Caption Generator Using DenseNet201 and ResNet50

Image Caption generation is an important research area in computer vision and natural language processing. This paper compares two popular Convolutional Neural Network (CNN) architectures, DenseNet201 and ResNet50, for feature extraction in the title generation task. The study aims to analyze the impact of these architectures on the quality of generated subtitles by measuring their learning curves and Bilingual Evaluation Understudy (BLEU) scores. The study shows that the choice of CNN architecture significantly affects the performance of the captioning model. Densenet201 and Resnet50 have different learning models and BLEU scores, indicating that the former is more effective at capturing high-level features, while the latter is more suitable for capturing local features. This study’s results will help develop more accurate and efficient subtitling models.

Accurate and robust auto-segmentation of head and neck organ-at-risks based on a novel CNN fine-tuning workflow.

Obvious inconsistencies in auto-segmentations exist among various AI software. In this study, we have developed a novel convolutional neural network (CNN) fine-tuning workflow to achieve precise and robust localized segmentation. The datasets include Hubei Cancer Hospital dataset, Cetuximab Head and Neck Public Dataset, and Québec Public Dataset. Seven organs-at-risks (OARs), including brain stem, left parotid gland, esophagus, left optic nerve, optic chiasm, mandible, and pharyngeal constrictor, were selected. The auto-segmentation results from four commercial AI software were first compared with the manual delineations. Then a new multi-scale lightweight residual CNN model with an attention module (named as HN-Net) was trained and tested on 40 samples and 10 samples from Hubei Cancer Hospital, respectively. To enhance the network's accuracy and generalization ability, the fine-tuning workflow utilized an uncertainty estimation method for automatic selection of candidate samples of worthiness from Cetuximab Head and Neck Public Dataset for further training. The segmentation performances were evaluated on the Hubei Cancer Hospital dataset and/or the entire Québec Public Dataset. A maximum difference of 0.13 and 0.7mm in average Dice value and Hausdorff distance value for the seven OARs were observed by four AI software. The proposed HN-Net achieved an average Dice value of 0.14 higher than that of the AI software, and it also outperformed other popular CNN models (HN-Net: 0.79, U-Net: 0.78, U-Net++: 0.78, U-Net-Multi-scale: 0.77, AI software: 0.65). Additionally, the HN-Net fine-tuning workflow by using the local datasets and external public datasets further improved the automatic segmentation with the average Dice value by 0.02. The delineations of commercial AI software need to be carefully reviewed, and localized further training is necessary for clinical practice. The proposed fine-tuning workflow could be feasibly adopted to implement an accurate and robust auto-segmentation model by using local datasets and external public datasets.

Transfer learning with CNNs for efficient prostate cancer and BPH detection in transrectal ultrasound images

Early detection of prostate cancer (PCa) and benign prostatic hyperplasia (BPH) is crucial for maintaining the health and well-being of aging male populations. This study aims to evaluate the performance of transfer learning with convolutional neural networks (CNNs) for efficient classification of PCa and BPH in transrectal ultrasound (TRUS) images. A retrospective experimental design was employed in this study, with 1380 TRUS images for PCa and 1530 for BPH. Seven state-of-the-art deep learning (DL) methods were employed as classifiers with transfer learning applied to popular CNN architectures. Performance indices, including sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), Kappa value, and Hindex (Youden’s index), were used to assess the feasibility and efficacy of the CNN methods. The CNN methods with transfer learning demonstrated a high classification performance for TRUS images, with all accuracy, specificity, sensitivity, PPV, NPV, Kappa, and Hindex values surpassing 0.9400. The optimal accuracy, sensitivity, and specificity reached 0.9987, 0.9980, and 0.9980, respectively, as evaluated using twofold cross-validation. The investigated CNN methods with transfer learning showcased their efficiency and ability for the classification of PCa and BPH in TRUS images. Notably, the EfficientNetV2 with transfer learning displayed a high degree of effectiveness in distinguishing between PCa and BPH, making it a promising tool for future diagnostic applications.

Thai culture image classification with transfer learning

Classifying images of Thai culture is important for a variety of applications, such as tourism, education, and cultural preservation. However, building a Machine learning model from scratch to classify Thai cultural images can be challenging due to the limited availability of annotated data. In this study, we investigate the use of transfer learning for the task of image classification on a dataset of Thai cultural images. We utilize three popular convolutional neural network models, namely MobileNet, EfficientNet, and residual network (ResNet) as baseline pre-trained models. Their performances were evaluated when they were trained from random initialization, used as a feature extractor, and fully fine-tuned. The results showed that all three models performed better in terms of accuracy and training time when they were used as a feature extractor, with EfficientNet achieving the highest accuracy of 95.87% while maintaining the training time of 24 ms/iteration. To better understand the reasoning behind the predictions made by the models, we deployed the gradient-weighted class activation mapping (Grad-CAM) visualization technique to generate heatmaps that the models attend to when making predictions. Both our quantitative and qualitative experiments demonstrated that transfer learning is an effective approach to image classification on Thai cultural images.

Attention driven CWT-deep learning approach for discrimination of Radar PRI modulation

With the proliferation of radio frequency (RF) systems and radar applications, Electronic Warfare (EW) is receiving increasing importance. The analysis of the radar signals is a critical EW task that decides the nature of counter employments. In an Electronic Support (ES) system, the challenge is to detect hostile radiation sources efficiently and trigger a counter response. Detection of types of Pulse Repetition Interval (PRI) modulation of radar signal significantly facilitates the manifestation of RF emitters during recognition which is difficult in a dense EW environment. Recent developments in artificial intelligence (AI) methods suggest that this emerging technology can be effective for such purposes. In this direction, an automatic approach for recognizing several kinds of complex PRI modulation based on Continuous Wavelet Transform (CWT) and a combination of the vanilla Convolutional Neural Network (CNN), a multi-head self-attention (MHSA) mechanism and the popular Long Short-Term Memory (LSTM) is proposed. The CWT is used to decompose the PRI modulation sequence and obtain different time–frequency components. Further, aided by the proposed CNN-MHSA-LSTM combination, the features extracted from the CWT 2D-scalograms are used to execute PRI modulation discrimination. In this method, the vanilla CNN is employed for the extraction of deep features to figure out the class details while capturing the spatial attributes. Thereafter, to improve the discriminative power of the entire framework a MHSA mechanism is used. The temporal attributes are acquired by the LSTM which works in concert with the CNN for executing the detection of the PRI classes based on the extracted features. Also to assess the effectiveness of the proposed method, three models based on ResNet, popular CNN and SqueezeNet are implemented for benchmark comparison in terms of overall performance and complexity. The simulation results show that the proposed method enhances performance and achieves robustness in the noise-filled and imperfect channel knowledge environment. The best recognition accuracy is 98.3% with 50% spurious pulses in the environment which fluctuates with imperfect channel knowledge cases.

FPGA-Based CNN for Eye Detection in an Iris Recognition at a Distance System

Neural networks are the state-of-the-art solution to image-processing tasks. Some of these neural networks are relatively simple, but the popular convolutional neural networks (CNNs) can consist of hundreds of layers. Unfortunately, the excellent recognition accuracy of CNNs comes at the cost of very high computational complexity, and one of the current challenges is managing the power, delay and physical size limitations of hardware solutions dedicated to accelerating their inference process. In this paper, we describe the embedding of an eye detection system on a Zynq XCZU4EV UltraScale+ multiprocessor system-on-chip (MPSoC). This eye detector is used in the application framework of a remote iris recognition system, which requires high resolution images captured at high speed as input. Given the high rate of eye regions detected per second, it is also important that the detector only provides as output images eyes that are in focus, discarding all those seriously affected by defocus blur. In this proposal, the network will be trained only with correctly focused eye images to assess whether it can differentiate this pattern from that associated with the out-of-focus eye image. Exploiting the neural network’s advantage of being able to work with multi-channel input, the inputs to the CNN will be the grey level image and a high-pass filtered version, typically used to determine whether the iris is in focus or not. The complete system synthetises other cores and implements CNN using the so-called Deep Learning Processor Unit (DPU), the intellectual property (IP) block released by AMD/Xilinx. Compared to previous hardware designs for implementing FPGA-based CNNs, the DPU IP supports extensive deep learning core functions, and developers can leverage DPUs to conveniently accelerate CNN inference. Experimental validation has been successfully addressed in a real-world scenario working with walking subjects, demonstrating that it is possible to detect only eye images that are in focus. This prototype module includes a CMOS digital image sensor that provides 16 Mpixel images, and outputs a stream of detected eyes as 640 × 480 images. The module correctly discards up to 95% of the eyes present in the input images as not being correctly focused.

Reduced-complexity Convolutional Neural Network in the compressed domain

Deep neural networks have achieved outstanding performance in computer vision tasks. Convolutional Neural Networks (CNNs) typically operate in the spatial domain with raw images, but in practice, images are usually stored and transmitted in their compressed representation where JPEG is one of the most widely used encoder. Also, these networks are computationally intensive and slow. This paper proposes performing the learning and inference processes in the compressed domain in order to reduce the computational complexity and improve the speed of popular CNNs. For this purpose, a novel graph-based frequency channel selection method is proposed to identify and select the most important frequency channels. The computational complexity is reduced by retaining the important frequency components and discarding the insignificant ones as well as eliminating the unnecessary layers of the network. Experimental results show that the modified ResNet-50 operating in the compressed domain is up to 70% faster than the spatial-based traditional ResNet-50 while resulting in similar classification accuracy. Moreover, this paper proposes a preprocessing step with partial encoding to improve the resilience to distortions caused by low-quality encoded images. Finally, we show that training a network with highly compressed data can achieve a good classification accuracy with up to 93% reduction in the storage requirements of the training data.

An attentional residual feature fusion mechanism for sheep face recognition

In the era of globalization and digitization of livestock markets, sheep are considered an essential source of food production worldwide. However, sheep behavior monitoring, disease prevention, and precise management pose urgent challenges in the development of smart ranches. To address these problems, individual identification of sheep has become an increasingly viable solution. Despite the benefits of traditional sheep individual identification methods, such as accurate tracking and record-keeping, they are labor-intensive and inefficient. Popular convolutional neural networks (CNNs) are unable to extract features for specific problems, further complicating the issue. To overcome these limitations, an Attention Residual Module (ARM) is proposed to aggregate the feature mapping between different layers of the CNN. This approach enables the general model of the CNN to be more adaptable to task-specific feature extraction. Additionally, a targeted sheep face recognition dataset containing 4490 images of 38 individual sheep has been constructed. Furthermore, the experimental data was expanded using image enhancement techniques such as rotation and panning. The results of the experiments indicate that the accuracy of the VGG16, GoogLeNet, and ResNet50 networks with the ARM improved by 10.2%, 6.65%, and 4.38%, respectively, compared to these recognition networks without the ARM. Therefore, the proposed method for specific sheep face recognition tasks has been proven effective.

A Deep Learning Framework for Intelligent Fault Diagnosis Using AutoML-CNN and Image-like Data Fusion

Intelligent fault diagnosis (IFD) is essential for preventative maintenance (PM) in Industry 4.0. Data-driven approaches have been widely accepted for IFD in smart manufacturing, and various deep learning (DL) models have been developed for different datasets and scenarios. However, an automatic and unified DL framework for developing IFD applications is still required. Hence, this work proposes an efficient framework integrating popular convolutional neural networks (CNNs) for IFD based on time-series data by leveraging automated machine learning (AutoML) and image-like data fusion. After normalisation, uniaxial or triaxial signals are reconstructed into -channel pseudo-images to satisfy the input requirements for CNNs and achieve data-level fusion simultaneously. Then, the model training, hyperparameter optimisation, and evaluation can be taken automatically based on AutoML. Finally, the selected model can be deployed on a cloud server or an edge device (via tiny machine learning). The proposed framework and method were validated via two case studies, demonstrating the framework’s availability for the automatic development of IFD applications and the effectiveness of the proposed data-level fusion method.

Hybrid morphological-convolutional neural networks for computer-aided diagnosis.

Training deep Convolutional Neural Networks (CNNs) presents challenges in terms of memory requirements and computational resources, often resulting in issues such as model overfitting and lack of generalization. These challenges can only be mitigated by using an excessive number of training images. However, medical image datasets commonly suffer from data scarcity due to the complexities involved in their acquisition, preparation, and curation. To address this issue, we propose a compact and hybrid machine learning architecture based on the Morphological and Convolutional Neural Network (MCNN), followed by a Random Forest classifier. Unlike deep CNN architectures, the MCNN was specifically designed to achieve effective performance with medical image datasets limited to a few hundred samples. It incorporates various morphological operations into a single layer and uses independent neural networks to extract information from each signal channel. The final classification is obtained by utilizing a Random Forest classifier on the outputs of the last neural network layer. We compare the classification performance of our proposed method with three popular deep CNN architectures (ResNet-18, ShuffleNet-V2, and MobileNet-V2) using two training approaches: full training and transfer learning. The evaluation was conducted on two distinct medical image datasets: the ISIC dataset for melanoma classification and the ORIGA dataset for glaucoma classification. Results demonstrate that the MCNN method exhibits reliable performance in melanoma classification, achieving an AUC of 0.94 (95% CI: 0.91 to 0.97), outperforming the popular CNN architectures. For the glaucoma dataset, the MCNN achieved an AUC of 0.65 (95% CI: 0.53 to 0.74), which was similar to the performance of the popular CNN architectures. This study contributes to the understanding of mathematical morphology in shallow neural networks for medical image classification and highlights the potential of hybrid architectures in effectively learning from medical image datasets that are limited by a small number of case samples.

Enhancing Brain Tumor Detection on MRI Images Using an Innovative VGG-19 Model-Based Approach

Early detection and diagnosis of brain tumors have a critical impact on the treatment of brain tumor patients. This is because initiating interventions early directly impacts the patient's chances of continuing their life. In the field of medical research, various methods are employed for the detection of brain tumors. Among these methods, magnetic resonance imaging (MRI) is the most popular due to its superior image quality. By leveraging technological advancements, the utilization of deep learning techniques in the identification of brain tumors ensures both high accuracy and simplification of the process. In a conducted study, a new model was developed by utilizing the VGG-19 architecture, a popular convolutional neural network model, to achieve high accuracy in brain tumor detection. In the study, precision, F1 score, accuracy, specificity, Matthews correlation coefficient, and recall metrics were used to evaluate the performance of the developed model. The deep learning model developed for brain tumor detection was trained and evaluated on an open-source dataset consisting of MRI images of gliomas, meningiomas, pituitary tumors, and healthy brains. The results obtained from the study demonstrate the promising potential of using the developed model in clinical applications for brain tumor detection. The high accuracy achieved by the developed model emphasizes its potential as an auxiliary resource for healthcare professionals in brain tumor detection. This research aims to evaluate the model as a valuable tool that can assist physicians in making informed treatment decisions regarding brain tumor diagnosis.

Image-Based Lunar Hazard Detection in Low Illumination Simulated Conditions via Vision Transformers.

Hazard detection is fundamental for a safe lunar landing. State-of-the-art autonomous lunar hazard detection relies on 2D image-based and 3D Lidar systems. The lunar south pole is challenging for vision-based methods. The low sun inclination and the terrain rich in topographic features create large areas in shadow, hiding the terrain features. The proposed method utilizes a vision transformer (ViT) model, which is a deep learning architecture based on the transformer blocks used in natural language processing, to solve this problem. Our goal is to train the ViT model to extract terrain features information from low-light RGB images. The results show good performances, especially at high altitudes, beating the UNet, one of the most popular convolutional neural networks, in every scenario.

Comparing torsional and lateral vibration data for deep learning-based drive train gear diagnosis

Deep learning-based fault diagnosis models have been demonstrated to recognise machine health conditions from vibration data. However, most related studies have focused on lateral vibration data, and mostly neglected torsional vibration data. Yet, torsional vibration data can provide an advantage in diagnosing gear faults. Torsional vibration is typically less noisy than lateral vibration data as it can be measured directly from the rotating components. To this end, this study presents a large gear fault dataset with artificial faults of varying severity acquired from a downscaled thruster test rig operated at a vast range of rotating speeds. The test rig was equipped with multiple torque transducers, rotary encoders and piezoelectric accelerometers. The lateral and torsional vibration data acquired with these sensors were evaluated with three popular convolutional neural networks in extensive ablation studies. An interpretability analysis was conducted based on amplitude spectra and Grad-CAM visualisations. The results demonstrate that torsional vibration can be an effective source of data for gear fault diagnosis. For example, the models diagnose the most difficult gear conditions using only one torque transducer more accurately than using three accelerometers mounted on the gear box. Furthermore, the highest accuracy in each ablation study related to experiments with combined lateral and torsional vibration data. In addition, the interpretability analysis showed that the lower frequencies had relatively higher amplitudes in torsional vibration than in lateral vibration. The interpretability analysis also indicates that the models reached higher classification accuracies with torsional vibration data due to the lower dominating frequencies. Overall, this study highlights the potential benefits of using torsional vibration data for deep learning-based fault diagnosis of gears.

Galaxy morphology classification using VGG16

There are billions of galaxies in the universe, each with millions or billions of stars. Galaxies vary in shape and size, and they interact with one another, sometimes colliding and merging, sometimes tearing apart, resulting in different features such as bars, rings, pseudo-pillars, and so on. These morphological properties play an important role in the study of galaxy evolution, so the classification of galaxy morphology has always been crucial in astrophysics. Earlier, using traditional methods or manual inspection to accomplish the classification task was a great challenge for astronomers. Today, however, machine learning is widely used for galaxy morphology classification because of its evolving efficiency and accuracy. In this paper, a popular convolutional neural network model (VGG16) will be used to classify galaxy morphology with an accuracy of 81%, which means that VGG16 can be used for correct classification.

Comprehensive Evaluation of Deep Neural Network Architectures for Parawood Pith Estimation

Accurate pith estimation is crucial for maintaining the quality of wood products. This study delves into deep learning techniques for precise Parawood pith estimation, employing popular convolutional neural networks (ResNet50, MobileNet, and Xception) with adapted regression heads. Through variations in regression functions, optimizers, and training epochs, the most effective models were pinpointed. Xception, coupled with Huber Loss regression, Nadam optimizer, and 200 epochs, showcased superior performance, achieving a 4.48 mm mean error (with a standard deviation of 3.69 mm) in Parawood. Notably, benchmarking on the Douglas Fir dataset yielded similar results (2.81 mm mean error, standard deviation: 1.57 mm). These findings underscore deep learning's potential for Parawood and Douglas Fir pith estimation, offering substantial benefits to wood industry quality control and production efficiency. By harnessing advanced machine learning techniques, this study advances wood industry processes, promoting the adoption of state-of-the-art technology in forestry and wood science. Doi: 10.28991/HIJ-2023-04-03-06 Full Text: PDF