Basic Convolutional Neural Network Research Articles

Noise2Variance: Dual networks with variance constraint for self‐supervised real‐world image denoising

AbstractImage denoising aims to restore a clean image from a noisy image. Traditional methods utilizing convolutional neural networks (CNN) for denoising are trained using pairs of noisy and clean images to comprehend the transformation from a noisy image to a clean one. However, the acquisition of such image pairs in real‐world scenarios presents a challenge. Hence, numerous self‐supervised denoising techniques have been developed that do not require clean images for training. This study demonstrates that a straightforward loss design, concentrating on variance, can effectively train a standard CNN denoiser in a self‐supervised fashion. A novel theoretical framework is introduced for training a basic CNN denoising model using three constraints: mean, variance, and augmentation. The variance constraint is crucial as it prevents the trained model from converging to trivial solutions such as identity or zero mapping. This theory provides valuable insights for the development of new self‐supervised denoising methods. Furthermore, a method that applies this theory to proposed dual networks is developed, which consist of two standard CNN models predicting both the clean image and the noise. This approach enhances model capacity during training while minimizing computational costs during inference. This method exemplifies the implementation of the variance constraint and introduces a data constraint for dual networks. Notably, the proposed method only assumes the presence of additive white noise, irrespective of the noise distribution. This minimal assumption enhances the model's robustness against noise with complex or unknown distributions in real‐world distorted images. Experimental results indicate that the proposed Noise2Variance method exhibits commendable performance on peak signal noise ratio and structural similarity metrics compared to existing self‐supervised denoising techniques. Visual comparison of results further substantiates the efficacy of the proposed method. A comparison of model complexity reveals that the method is efficient among the compared CNN‐based techniques.

Decoding Innocence: Advancing Forensic Facial Discrimination through Comparative Analysis of Conventional CNN and Advanced Architectures

The field of Digital Image Forensics (DIF) faces a critical issue in accurately identifying children in digital images, notably in cases involving the proliferation of child sexual abuse content. Existing techniques face hurdles due to model architecture limitations, dataset suitability concerns, and classification imbalance, impeding their ability to recognize children to deter pornographic images. Addressing this challenge, this study introduces Implicit Feature Extraction (IFE), a specialized approach for distinguishing child and adult images in object detection. Leveraging Convolutional Neural Networks (CNNs), the IFE method automates the extraction of discriminative facial features, surpassing the constraints of Explicit Feature Extraction (EFE) methods, which achieve an accuracy of around 70%. The research focuses on three core objectives introducing IFE for detailed face feature detection in DIF's child and adult image identification, implementing IFE with CNNs to enhance image classification, and conducting a thorough evaluation of the proposed technique's performance using key metrics like accuracy and balanced classification results and comparing the result with a basic CNN model’s performance. This research's significance lies in its notable contributions to digital image forensics, particularly in combatting child exploitation. The fusion of IFE with CNNs showcases 92% accuracy in distinguishing child and adult images, promising advancements with practical implications in child protection and forensic investigations. The comprehensive evaluation using the UTKFace dataset underscores the proposed technique's efficacy, marking a substantial improvement in child image identification within digital image forensics.

A deep learning approach for Maize Lethal Necrosis and Maize Streak Virus disease detection

Maize is an important crop cultivated in Sub-Saharan Africa, essential for food security. However, its cultivation faces significant challenges due to debilitating diseases such as Maize Lethal Necrosis (MLN) and Maize Streak Virus (MSV), which can lead to severe yield losses. Traditional plant disease diagnosis methods are often time-consuming and prone to errors, necessitating more efficient approaches. This study explores the application of deep learning, specifically Convolutional Neural Networks (CNNs), in the automatic detection and classification of maize diseases. We investigate six architectures: Basic CNN, EfficientNet V2 B0 and B1, LeNet-5, VGG-16, and ResNet50, using a dataset of 15344 images comprising MSV, MLN, and healthy maize leaves. Additionally, We performed hyperparameter tuning to improve the performance of the models and Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability. Our results show that the EfficientNet V2 B0 model demonstrated an accuracy of 99.99% in distinguishing between healthy and disease-infected plants. The results of this study contribute to the advancement of AI applications in agriculture, particularly in diagnosing maize diseases within Sub-Saharan Africa.

Resolving data imbalance in alkaline battery defect detection: a voting-based deep learning approach

Alkaline battery defect detection is crucial for ensuring product quality and providing diagnostic feedback. Recently, high-performance deep learning algorithms have been introduced to recognise defects in alkaline batteries. However, the majority of deep learning-based methods overlook the significant data imbalance issues in alkaline battery electrode images, potentially resulting in performance degradation. Therefore, a voting-based recognition algorithm containing three parts is proposed in this study. Firstly, a resampling and training method is developed to provide richer information and stronger constraints. Secondly, a weak classification framework based on an improved convolutional neural network (CNN) is designed to provide fine-grained category representations. Finally, a voting-based prediction approach is proposed to improve accuracy and obtain the final results. Visualisation results demonstrate that the proposed algorithm has a stronger clustering ability and uses voting-based prediction to improve performance. Comparison research shows that the proposed method significantly improves the recall of minority classes and the precision of majority classes and reaches a state-of-the-art F1 score of 0.982 for alkaline battery defect recognition, which is 0.041 higher than the basic CNN model.

A multi-step loss meta-learning method based on multi-scale feature extraction for few-shot fault diagnosis

Existing deep learning (DL) algorithms are based on a large amount of training data and they face challenges in effectively extracting fault features when dealing with few-shot fault diagnoses. Model-agnostic meta-learning (MAML) also faces some challenges, including the limited capability of the basic convolutional neural network (CNN) with a single convolutional kernel to extract fault features comprehensively, as well as the instability of model training due to the inner and outer double-layer loops. To address these issues, this paper presents a multi-step loss meta-learning method based on multi-scale feature extraction (MFEML). Firstly, an improved multi-scale feature extraction module (IMFEM) is designed to solve the problem of the insufficient feature extraction capability of the CNN. Secondly, the multi-step loss is used to reconstruct the meta-loss to address the issue of MAML training instability. Finally, the experimental results of two datasets demonstrate the effectiveness of the MFEML.

A Global Seawater Density Distribution Model Using a Convolutional Neural Network.

Seawater density is an important physical property in oceanography that affects the accuracy of calculations such as gravity fields and tidal potentials and the calibration of acoustic and optical oceanographic sensors. In related studies, constant density values are frequently used, which can introduce significant errors. Therefore, this study employs a basic convolutional neural network model to construct a comprehensive model showing the seawater density distribution across the globe. The model takes into account depth, latitude, longitude, and month as inputs. Numerous real seawater datasets were used to train the model, and it has been shown that the model has an absolute mean error and root mean square error of less than 1 kg/m3 in 99% of the test set samples. The model effectively demonstrates the influence of input parameters on the distribution of seawater density. In this paper, we present a newly developed global model for distributing seawater density which is both comprehensive and accurate, surpassing previous models. The utilization of the model presented in this paper for estimating seawater density can minimize errors in theoretical ocean models and serve as a foundation for designing and analyzing ocean exploration systems.

Image Forgery Detection Based on Parallel Convolutional Neural Networks

Abstract— Due to the availability of deep networks, progress has been made in the field of image recognition. Images and videos are spreading very conveniently and with the availability of strong editing tools the tampering of digital content become easy. To detect such scams, we proposed techniques. In our paper, we proposed two important aspects of employing deep convolutional neural networks to image forgery detection. We first explore and examine different preprocessing method along with convolutional neural networks (CNN) architecture. Later we evaluated the different transfer learning for pre-trained ImageNet(via-fine-tuning) and implement it over our dataset CASIA V2.0. So, it covers the pre-processing techniques with basic CNN model and later see the powerful effect of the transfer learning models. Keywords— image tampering, convolution neural network (CNN), error level analysis (ELA), transfer learning, sharpening filter, fine-tuning

Genetic Algorithm-Based Data Optimization for Efficient Transfer Learning in Convolutional Neural Networks: A Brain–Machine Interface Implementation

In brain–machine interface (BMI) systems, the performance of trained Convolutional Neural Networks (CNNs) is significantly influenced by the quality of the training data. Another issue is the training time of CNNs. This paper introduces a novel approach by combining transfer learning and a Genetic Algorithm (GA) to optimize the training data of CNNs. Transfer learning is implemented across different subjects, and the data chosen by GA aim to improve CNN performance. In addition, the GA-selected data shed light on the similarity in brain activity between subjects. Two datasets are used: (1) the publicly available BCI Competition IV, in which the subjects performed motor imagery (MI) tasks, and (2) the dataset created by healthy subjects of our laboratory performing motor movement (MO) tasks. The experimental results indicate that the brain data selected by the GA improve the recognition accuracy of the target CNN (TCNN) using pre-trained base CNN (BCNN). The improvement in accuracy is 11% and 4% for the BCI Competition IV and our laboratory datasets, respectively. In addition, the GA-selected training data reduce the CNN training time. The performance of the trained CNN, utilizing transfer learning, is tested for real-time control of a robot manipulator.

IMCNN:Intelligent Malware Classification using Deep Convolution Neural Networks as Transfer learning and ensemble learning in honeypot enabled organizational network

Traditional malware detection systems based on signature-based detection methods cannot detect new and unseen malware. Moreover, conventional machine learning methods for malware detection have utilized features extracted through static program analysis or dynamic analysis, which requires code debugging and execution primarily through offline processing; hence not a scalable approach. This paper proposes a novel intelligent malware classification using a deep convolution neural network (IMCNN) in organizational networks enabled with Honeypots. Systematic customization of pre-trained convolutional neural networks(CNN) as a transfer learning and ensemble learning as a classification is presented to detect intelligent modern-day malware. Real-world malware samples are systematically labeled and visualized into grayscale images. Four cutting-edge deep CNN models - VGG16, VGG19, InceptionV3, and ResNet50, are trained on the ImageNet database (≥1 million) and fine-tuned as feature extractors along with a basic CNN model. Three strategies are designed for feature extraction and selection: Rectified linear unit (ReLU) fully connected layer embedded in a deep CNN model, principal component analysis (PCA), and singular value decomposition(SVD). Reduced sets of features are stacked and used to train k-nearest neighbor (k-NN), support vector machine (SVM), and random forest (RF) classifiers for predictions. Subsequently, the predictive probabilities of different machine-learned models are ensembled using a soft voting method for final classification. The proposed method is evaluated on MalImg datasets (9339 malware samples of 25 families) and real-world modern malware datasets (690 malware of 22 families). The experimental results reveal that despite using a reduced feature set, the IMCNN effectively detects malware with 99.36% test accuracy on unseen data for MalImg datasets and 92.11% for real-world malware. In addition, the proposed method is compared with several existing state-of-art malware detection models in terms of performance accuracy and found performing as the best. Experiments demonstrated that the proposed method is resilient to polymorphic code obfuscation used by the malware authors.

Deep learning based Non-Local k-best renyi entropy for classification of white blood cell subtypes

The major issue in classifying blood cells from the microscopic image is overlapping cells and overfitting data. Therefore, to solve these problems in microscopic imaging, a thresholding scheme based on a non-local average filter approach is used to segment cells that separate the overlap of cells and thus perform the segmentation procedure. Basic convolutional neural networks (CNN) such as AlexNet, ResNet, and DenseNet have proven to be more powerful in image classification in various fields. However, the problem is that the number of parameters used in the network increased, which tends to separate the optimization procedure; otherwise, it leads to complexity in processing (increased simulation time). The SqueezeNet architecture is used to reduce these problems and overcome the problems in blood cell classification, like the overfitting of data and gradient vanishing problems. That will extract and classify the features based on the ground truth with 5 times fewer parameters than the traditional CNN architectures. The proposed method has improved the accuracy to 99.83%, 99.67%, 99.56% and 99.02% for the test datasets of kaggle-blood cell image, LISC dataset, Raabin-WBC and BCCD dataset, respectively.

A comparative analysis of converters of tabular data into image for the classification of Arboviruses using Convolutional Neural Networks

Tabular data is commonly used in business and literature and can be analyzed using tree-based Machine Learning (ML) algorithms to extract meaningful information. Deep Learning (DL) excels in data such as image, sound, and text, but it is less frequently utilized with tabular data. However, it is possible to use tools to convert tabular data into images for use with Convolutional Neural Networks (CNNs) which are powerful DL models for image classification. The goal of this work is to compare the performance of converters for tabular data into images, select the best one, optimize a CNN using random search, and compare it with an optimized ML algorithm, the XGBoost. Results show that even a basic CNN, with only 1 convolutional layer, can reach comparable metrics to the XGBoost, which was trained on the original tabular data and optimized with grid search and feature selection. However, further optimization of the CNN with random search did not significantly improve its performance.

Deep Mutual Learning-Based Mode Recognition of Orbital Angular Momentum

Due to its orbital angular momentum (OAM), optical vortex has been widely used in communications and LIDAR target detection. The OAM mode recognition based on deep learning is mostly based on the basic convolutional neural network. To ensure high-precision OAM state detection, a deeper network structure is required to overcome the problem of similar light intensity distribution of different superimposed vortex beams and the effect of atmospheric turbulence disturbance. However, the large number of parameters and the computation of the OAM state detection network conflict with the requirements of deploying optical communication system equipment. In this paper, an online knowledge distillation scheme is selected to achieve an end-to-end single-stage training and the inter-class dark knowledge of similar modes are fully utilized. An optical vortex OAM state detection technique based on deep mutual learning (DML) is proposed. The simulation results show that after mutual learning training, a small detection network with higher accuracy can be obtained, which is more suitable for terminal deployment. Based on the scalability of the number of networks in the DML queue, it provides a new possibility to further improve the detection accuracy of the optical communication.

CNN combined with attention module for facial expression recognition

In contemporary times, the proliferation of information technology has prompted an increased interest among scholars towards artificial intelligence, particularly its constituent algorithms, such as deep learning, multilayer perceptron and convolutional neural networks. Of specific interest is the analysis of facial expressions, which has emerged as a popular research topic. However, classifying facial expressions presents a significant challenge due to variations in expressions associated with different emotions, as well as similarities between the various emotions. The task of expression classification is further complicated by the abundance of facial features that must be considered. In this paper, the basic Convolutional Neural Network (CNN) model was first trained on the FRE-2013 face expression dataset and optimized by adding an attention module to learn and analyse the key features of faces, thereby improving the classification accuracy of the model. In short, this model achieved an accuracy of 71.65% and an F1-Score of 0.66 while the model without the attention module only achieved a model accuracy of 69.86% and an F1-Score of 0.65, an improvement of about 2%. The analysis demonstrated that by adding the attention mechanism, some important features e.g. the eyes and mouth, are given more weight, thus improving the classification accuracy.

Phenotyping and predicting wheat spike characteristics using image analysis and machine learning

AbstractImprovements in trait phenotyping are needed to increase the quantity and quality of data available for genetic improvement of crops. In this study, we used moderate throughput image analysis and machine learning as a pipeline for phenotyping a key wheat spike characteristic: spikelet number per spike. A population of 594 soft red winter wheat inbred lines was evaluated in the field for 2 years and images of wheat spikes were taken and used to train deep‐learning algorithms to predict spikelet number. A total of 12,717 images were used to train, test, and validate a basic regression convolutional neural network (CNN), a visual geometry group application regression model, VGG16, the ResNet152V2 model, and the EfficientNetV2L model. The EfficientNetV2L model was the most accurate, having the lowest mean absolute error, second lowest root mean square error, and highest coefficient of determination (mean absolute error [MAE] = 0.60, root mean square error [RMSE] = 0.79, and R2 = 0.90). The ResNet152V2 model was slightly less accurate with a slightly better fit (MAE = 0.61,m RMSE = 0.78, and R2 = 0.87), followed by the basic CNN (MAE = 0.75, RMSE = 1.00, and R2 = 0.74) and finally by the VGG16 (MAE = 1.51, RMSE = 1.29, and R2 = 0.076). With an average error of just above one half of a spikelet, utilizing image analysis and machine learning counting methods could be used for multiple breeding applications, including direct selection of spikelet number, to provide data to identify quantitative trait loci, or for training whole genome selection models.

Oral cancer detection using convolutional neural network optimized by combined seagull optimization algorithm

Early diagnosis of oral cancer is crucial for improving patient outcomes and saving lives. However, inaccurate and improper diagnosis can hinder effective treatment. This paper presents a novel method for detecting oral cancer using an optimized version of Convolutional Neural Network (CNN). While basic CNNs have been widely used for image classification tasks, the incorporation of the Seagull Optimization Algorithm and Particle Swarm Optimization Algorithm in optimizing the CNN architecture specifically for oral cancer detection is a unique approach that is provided in this study. By combining these algorithms, the proposed method optimizes the CNN's architecture, parameters, and training process specifically for oral cancer detection. This optimization enhances the performance and accuracy of the CNN in identifying cancerous regions in oral images. Unlike previous approaches, our method incorporates advanced image processing techniques, including noise reduction, contrast enhancement, and data augmentation, to enhance the quality of input data extracted from the Oral Cancer (Lips and Tongue) images (OCI) dataset. The optimized CNN architecture uses its ability to learn intricate patterns and features from the enhanced images, enabling more accurate identification of cancerous regions. To evaluate the effectiveness of our approach, we compare it against Textural analysis, FCM, CNN, R-CNN, and ResNet-101 using four measurement indices. Results demonstrate that our proposed CSOA-based CNN system achieves the highest accuracy rate (96.94%) compared to other methods, indicating its superior performance in oral cancer detection. Furthermore, our precision rate of 94.65% and recall rate of 91.60% highlight the model's high correctness and positive classification ability. Finally, our proposed method achieves the highest F1-score (88.55%), emphasizing its superiority over other comparative methods. Through our innovative integration of the Seagull Optimization Algorithm and Particle Swarm Optimization Algorithm with CNN, coupled with advanced image processing techniques, we provide a reliable and effective solution for early detection of oral cancer.

Supervised Single Channel Speech Enhancement Method Using UNET

This paper proposes an innovative single-channel supervised speech enhancement (SE) method based on UNET, a convolutional neural network (CNN) architecture that expands on a few changes in the basic CNN architecture. In the training phase, short-time Fourier transform (STFT) is exploited on the noisy time domain signal to build a noisy time-frequency domain signal which is called a complex noisy matrix. We take the real and imaginary parts of the complex noisy matrix and concatenate both of them to form the noisy concatenated matrix. We apply UNET to the noisy concatenated matrix for extracting speech components and train the CNN model. In the testing phase, the same procedure is applied to the noisy time-domain signal as in the training phase in order to construct another noisy concatenated matrix that can be tested using a pre-trained or saved model in order to construct an enhanced concatenated matrix. Finally, from the enhanced concatenated matrix, we separate both the imaginary and real parts to form an enhanced complex matrix. Magnitude and phase are then extracted from the newly created enhanced complex matrix. By using that magnitude and phase, the inverse STFT (ISTFT) can generate the enhanced speech signal. Utilizing the IEEE databases and various types of noise, including stationary and non-stationary noise, the proposed method is evaluated. Comparing the exploratory results of the proposed algorithm to the other five methods of STFT, sparse non-negative matrix factorization (SNMF), dual-tree complex wavelet transform (DTCWT)-SNMF, DTCWT-STFT-SNMF, STFT-convolutional denoising auto encoder (CDAE) and casual multi-head attention mechanism (CMAM) for speech enhancement, we determine that the proposed algorithm generally improves speech quality and intelligibility at all considered signal-to-noise ratios (SNRs). The suggested approach performs better than the other five competing algorithms in every evaluation metric.

Multi-Object Tracking in Heterogeneous environments (MOTHe) for animal video recordings.

Aerial imagery and video recordings of animals are used for many areas of research such as animal behaviour, behavioural neuroscience and field biology. Many automated methods are being developed to extract data from such high-resolution videos. Most of the available tools are developed for videos taken under idealised laboratory conditions. Therefore, the task of animal detection and tracking for videos taken in natural settings remains challenging due to heterogeneous environments. Methods that are useful for field conditions are often difficult to implement and thus remain inaccessible to empirical researchers. To address this gap, we present an open-source package called Multi-Object Tracking in Heterogeneous environments (MOTHe), a Python-based application that uses a basic convolutional neural network for object detection. MOTHe offers a graphical interface to automate the various steps related to animal tracking such as training data generation, animal detection in complex backgrounds and visually tracking animals in the videos. Users can also generate training data and train a new model which can be used for object detection tasks for a completely new dataset. MOTHe doesn't require any sophisticated infrastructure and can be run on basic desktop computing units. We demonstrate MOTHe on six video clips in varying background conditions. These videos are from two species in their natural habitat-wasp colonies on their nests (up to 12 individuals per colony) and antelope herds in four different habitats (up to 156 individuals in a herd). Using MOTHe, we are able to detect and track individuals in all these videos. MOTHe is available as an open-source GitHub repository with a detailed user guide and demonstrations at: https://github.com/tee-lab/MOTHe-GUI.

Performance Influencing Factors of Convolutional Neural Network Models for Classifying Certain Softwood Species

This study aims to verify the wood classification performance of convolutional neural networks (CNNs), such as VGG16, ResNet50, GoogLeNet, and basic CNN architectures, and to investigate the factors affecting classification performance. A dataset from 10 softwood species consisted of 200 cross-sectional micrographs each from the total part, earlywood, and latewood of each species. We used 80% and 20% of each dataset for training and testing, respectively. To improve the performance of the architectures, the dataset was augmented, and the differences in classification performance before and after augmentation were compared. The four architectures showed a high classification accuracy of over 90% between species, and the accuracy increased with increasing epochs. However, the starting points of the accuracy, loss, and training speed increments differed according to the architecture. The latewood dataset showed the highest accuracy. The epochs and augmented datasets also positively affected accuracy, whereas the total part and non-augmented datasets had a negative effect on accuracy. Additionally, the augmented dataset tended to derive stable results and reached a convergence point earlier. In the present study, an augmented latewood dataset was the most important factor affecting classification performance and should be used for training CNNs.

Automatic Diagnosis of Melanoma Based on EfficientNet and Patch Strategy

Melanoma is a fatal skin disease, and there are many challenging tasks in the detection of melanoma through neural network at this stage. We propose a new method for melanoma diagnosis based on EfficientNet and patch strategy. The diagnosis method has three stages of operation. First, Cyclegan is applied offline to synthesize the under-represented category samples from the over-represented category samples, and the original image and the synthesized samples are combined into a new training set to complete the conditional image synthesis task; second, we creatively propose the patch strategy and implement the patch algorithm, and apply the patch strategy offline to obtain the patch image of the newly merged training set; finally, the newly merged training set and the obtained patch image are sent to the classification network. Where, the model we proposed is composed of three parts: Basic Convolutional Neural Network, Auxiliary Convolutional Neural Network, and Fusion Convolutional Neural Network, and applies a weighted integration strategy. We evaluated the proposed method on the ISIC 2016 Skin Injury Challenge classification dataset. Experiments show that the patch strategy plays an important role in the field of melanoma classification, and the melanoma detection method proposed in this paper obtains an accuracy of 0.852 and an AUC value of 0.854 on the test set. This method can focus the attention of the classification network on the meaningful area of the skin lesion image through manual intervention, and can effectively solve the problem of category imbalance, thereby improving the performance of skin lesion classification.

Improved Convolutional Neural Network YOLOv5 for Underwater Target Detection Based on Autonomous Underwater Helicopter

To detect a desired underwater target quickly and precisely, a real-time sonar-based target detection system mounted on an autonomous underwater helicopter (AUH) using an improved convolutional neural network (CNN) is proposed in this paper. YOLOv5 is introduced as the basic CNN network because of its strength, lightweight and fast speed. Due to the turbidity and weak illumination of an undesirable underwater environment, some attention mechanisms are added, and the structure of YOLOv5 is optimized to improve the performance of the detector for sonar images with a 1–3% increment of mAP which can be up to 80.2% with an average speed of 0.025 s (40 FPS) in the embedded device. It has been verified both in the school tank and outdoor open water that the whole detection system mounted on AUH performs well and meets the requirements of real time and light weight using limited hardware.