CNN Architecture Research Articles

LBO-MPAM: Ladybug Beetle Optimization-based multilayer perceptron attention module for segmenting the skin lesion and automatic localization

ABSTRACT In recent years, skin cancer has been the most dangerous disease noticed among people worldwide. Skin cancer should be identified earlier to reduce the rate of mortality. Employing dermoscopic images can identify and categorise skin cancer effectively. But, the visual evaluation is a complex procedure to be done in the dermoscopic image. However, Deep learning (DL) is an efficient method for skin cancer detection; however, segmenting the skin lesion and automatic localisation in an earlier stage is complicated. In this paper, a novel Ladybug Beetle Optimization-Double Attention Based Multilevel 1-D CNN (LBO-DAM 1-D CNN) technique is proposed to detect and classify skin cancer. To improve skin lesion type discriminability, the two types of attention modules are introduced. The Ultra-Lightweight Subspace Attention Module (ULSAM) is utilised for classifying the feature maps into different stages to validate the frequency from different image samples. However, the multilayer perceptron attention module (MLPAM) is determined to provide information regarding skin cancer classification and diminish the noise and unwanted data. To minimise data loss, it is then combined with hierarchical complementarity during classification. Second, a modified MLPAM is used to extract significant feature spaces for network learning, select the most important information, and reduce feature space redundancy. The Ladybug Beetle Optimization (LBO) algorithm provides the optimal classification solution by minimising the loss rate of DAM 1-D CNN architecture. The experimentation is conducted on three different datasets such as ISIC2020, HAM10000, and the melanoma detection dataset. The experimental results revealed that the proposed method is compared with different existing methods such as IMFO-KELM, Mask RCNN, M-SVM, DCNN-9, and TL-CNN with different datasets. These methods attained 94.56, 92.65, 90.56, 88.65, and 95.5 for the ISIC2020 dataset but the proposed method enhanced the classification performance by attaining 97.02. Also, the validation is based on metrics, namely, accuracy, precision, sensitivity, and F1-score of 97.03%, 97.05%, 97.58%, and 97.27% for a total of 500 epochs.

HmsU-Net: A hybrid multi-scale U-net based on a CNN and transformer for medical image segmentation

Accurate medical image segmentation is of great significance for subsequent diagnosis and analysis. The acquisition of multi-scale information plays an important role in segmenting regions of interest of different sizes. With the emergence of Transformers, numerous networks adopted hybrid structures incorporating Transformers and CNNs to learn multi-scale information. However, the majority of research has focused on the design and composition of CNN and Transformer structures, neglecting the inconsistencies in feature learning between Transformer and CNN. This oversight has resulted in the hybrid network's performance not being fully realized. In this work, we proposed a novel hybrid multi-scale segmentation network named HmsU-Net, which effectively fused multi-scale features. Specifically, HmsU-Net employed a parallel design incorporating both CNN and Transformer architectures. To address the inconsistency in feature learning between CNN and Transformer within the same stage, we proposed the multi-scale feature fusion module. For feature fusion across different stages, we introduced the cross-attention module. Comprehensive experiments conducted on various datasets demonstrate that our approach surpasses current state-of-the-art methods.

Enhancing pulmonary abnormality detection with an optimized CNN architecture incorporating depth-wise separable convolution and inception module

Enhancing pulmonary abnormality detection with an optimized CNN architecture incorporating depth-wise separable convolution and inception module

ERTNet: an interpretable transformer-based framework for EEG emotion recognition.

Emotion recognition using EEG signals enables clinicians to assess patients' emotional states with precision and immediacy. However, the complexity of EEG signal data poses challenges for traditional recognition methods. Deep learning techniques effectively capture the nuanced emotional cues within these signals by leveraging extensive data. Nonetheless, most deep learning techniques lack interpretability while maintaining accuracy. We developed an interpretable end-to-end EEG emotion recognition framework rooted in the hybrid CNN and transformer architecture. Specifically, temporal convolution isolates salient information from EEG signals while filtering out potential high-frequency noise. Spatial convolution discerns the topological connections between channels. Subsequently, the transformer module processes the feature maps to integrate high-level spatiotemporal features, enabling the identification of the prevailing emotional state. Experiments' results demonstrated that our model excels in diverse emotion classification, achieving an accuracy of 74.23% ± 2.59% on the dimensional model (DEAP) and 67.17% ± 1.70% on the discrete model (SEED-V). These results surpass the performances of both CNN and LSTM-based counterparts. Through interpretive analysis, we ascertained that the beta and gamma bands in the EEG signals exert the most significant impact on emotion recognition performance. Notably, our model can independently tailor a Gaussian-like convolution kernel, effectively filtering high-frequency noise from the input EEG data. Given its robust performance and interpretative capabilities, our proposed framework is a promising tool for EEG-driven emotion brain-computer interface.

On-the-fly image-level oversampling for imbalanced datasets of manufacturing defects

Visual defect recognition and its manufacturing applications have been an upcoming topic in recent AI research. Defect datasets are often severely imbalanced and can be additionally burdened with separating classes of high visual similarity. Although various methods of data augmentation have been proposed to mitigate the class imbalance, they often fail to cope with tinier minority classes or have fidelity issues with smaller defects while, at the same time, needing significant computational resources to train. Also, augmentation based on vector-based oversampling struggles to produce high-fidelity inputs and is hard to apply on custom CNN architectures, which often perform better for this type of problem. Our work presents an image-level oversampling method based on an instance-based image generator that can be applied to any CNN directly during the training process without increasing the order of training time required. It is based on identifying a small number of the most uncertain base samples close to the estimated class boundaries and using them as seeds for augmentation. The resulting images are of high visual quality preserving small class differences, and they also improve the classifier boundary leading to higher recall scores than other state-of-the-art approaches.

Implementasi Convolutional Neural Network Pada Robot Tikus Micromouse Untuk Navigasi Labirin Dengan Pendeteksi Garis Menggunakan Raspberry Pi

Micromouse is a science and technology project competition that began in the late 1970s. Simply put, the micromouse competition is an event where robotic mice can navigate a 16x16 maze without human assistance. Convolutional Neural Network, also known as CNN, is an extension of the Artificial Neural Network (ANN) designed to process two-dimensional data and is often used for classification, recognition, and prediction tasks. In the conducted research, a robotic mouse was designed using Raspberry Pi to control servo motor movements and utilize a camera to detect maze obstacles, such as intersections, through image scanning. The robotic mouse was tested using a CNN architecture model on TensorFlow, assisted by a line detection algorithm. The testing results indicated that the CNN model architecture showed signs of overfitting (the model learned features too well from the training data) with an accuracy of 96.53%, and the model accuracy evaluation result was 97.97% on the designed dataset. As part of the testing, a maze size of 5x5 was used. The testing of the Convolutional Neural Network model and line detection algorithm, when applied to the robotic mouse, demonstrated that the mouse could navigate according to the path. However, the recorded data was influenced by external light reflecting off the surface of the path, causing the line detection algorithm to perceive it as an obstacle, leading the robot to turn back..

Automatic detection of nostril and key markers in images

In the medical field, one of complex challenges in surgery involving the skin, cartilage, mucosa, and the skeletal platform is the cleft nasal deformity. The causes of deformity are nasal formed from improper fusion of the medial and lateral nasal prominences with the maxillary prominence during embryologic development. Primary rhinoplasty or first time of cleft nose repair is the nasal surgery for repairing the cleft nasal deformity. In post cheiloplasty, the primary rhinoplasty patients must use nasal splint for supporting surgical wound. However, regular nasal splints are expensive and only come with default nasal size. This limits nasal splint from supporting specific patient cases such as children. The goal of this study is to implement a program that can automatically detect the nostril and green marker in patient images for creating custom-made 3D nasal splints. The proposed method is created utilizing the CNN model. YOLO architecture is utilized because it is the one of several CNN architectures that works well with the face recognition task. The YOLOV5, YOLOV8, and YOLO-NAS are attempted and compared in the training phase. The model with a highest performance is selected, and fine-tuned for adjusting to be compatible with a patient dataset. The fine-tuned YOLOV8 reaches a mAP with 99.5%. The predicted images from fine-tuned model are used to perform body part segmentation like Otsu's thresholding and discover contour to locate essential features like the green marker and nostril in ellipse and bounding boxes. The nostril distance is calculated using the bounding box and ellipse. The columella distance or nostril gap is measured from the distance between nostrils in YOLO predicted label. Both distances are converted into centimeter scale and evaluated with ground truth value in volunteer and patient cases from nasal expert for inspecting efficiency between bounding box, ellipse, and size of marker. The result proves that the proposed model can detect crucial features with mean absolute error 0.102. Furthermore, the proposed model indicates that marker size has no effect on detection and distance, according to the findings of the experiments.

DPXception: a lightweight CNN for image-based date palm species classification.

Date palm species classification is important for various agricultural and economic purposes, but it is challenging to perform based on images of date palms alone. Existing methods rely on fruit characteristics, which may not be always visible or present. In this study, we introduce a new dataset and a new model for image-based date palm species classification. Our dataset consists of 2358 images of four common and valuable date palm species (Barhi, Sukkari, Ikhlas, and Saqi), which we collected ourselves. We also applied data augmentation techniques to increase the size and diversity of our dataset. Our model, called DPXception (Date Palm Xception), is a lightweight and efficient CNN architecture that we trained and fine-tuned on our dataset. Unlike the original Xception model, our DPXception model utilizes only the first 100 layers of the Xception model for feature extraction (Adapted Xception), making it more lightweight and efficient. We also applied normalization prior to adapted Xception and reduced the model dimensionality by adding an extra global average pooling layer after feature extraction by adapted Xception. We compared the performance of our model with seven well-known models: Xception, ResNet50, ResNet50V2, InceptionV3, DenseNet201, EfficientNetB4, and EfficientNetV2-S. Our model achieved the highest accuracy (92.9%) and F1-score (93%) among the models, as well as the lowest inference time (0.0513 seconds). We also developed an Android smartphone application that uses our model to classify date palm species from images captured by the smartphone's camera in real time. To the best of our knowledge, this is the first work to provide a public dataset of date palm images and to demonstrate a robust and practical image-based date palm species classification method. This work will open new research directions for more advanced date palm analysis tasks such as gender classification and age estimation.

Decomposition iteration strategy for low-dose CT denoising.

In the medical field, computed tomography (CT) is a commonly used examination method, but the radiation generated increases the risk of illness in patients. Therefore, low-dose scanning schemes have attracted attention, in which noise reduction is essential. We propose a purposeful and interpretable decomposition iterative network (DISN) for low-dose CT denoising. This method aims to make the network design interpretable and improve the fidelity of details, rather than blindly designing or using deep CNN architecture. The experiment is trained and tested on multiple data sets. The results show that the DISN method can restore the low-dose CT image structure and improve the diagnostic performance when the image details are limited. Compared with other algorithms, DISN has better quantitative and visual performance, and has potential clinical application prospects.

Zero-exemplar deep continual learning for crop disease recognition: a study of total variation attention regularization in vision transformers.

With the increasing integration of AI technology in the food industry, deep learning has demonstrated its immense potential in the domain of plant disease image recognition. However, there remains a gap in research between models capable of continual learning of new diseases and addressing the inherent catastrophic forgetting issue in neural networks. This study aims to comprehensively evaluate various learning strategies based on advanced computer vision models for multi-disease continual learning tasks in food disease recognition. To cater to the benchmark dataset requirements, we collected the PlantDiseaseCL dataset, sourced from the internet, encompassing diverse crop diseases from apples, corn, and more. Utilizing the Vision Transformer (ViT) model, we established a plant disease image recognition classifier, which, in joint learning, outperformed several comparative CNN architectures in accuracy (0.9538), precision (0.9532), recall (0.9528), and F1 score (0.9560). To further harness the potential of ViT in food disease defect recognition, we introduced a mathematical paradigm for crop disease recognition continual learning. For the first time, we proposed a novel ViT-TV architecture in the multi-disease image recognition scenario, incorporating a Total Variation (TV) distance-based loss (TV-Loss) to quantify the disparity between current and previous attention distributions, fostering attention consistency and mitigating the catastrophic forgetting inherent in ViT without prior task samples. In the incremental learning of the PlantDiseaseCL dataset across 3-Steps and 5-Steps, our strategy achieved average accuracies of 0.7077 and 0.5661, respectively, surpassing all compared Zero-Exemplar Approaches like LUCIR, SI, MAS, and even outperforming exemplar-based strategies like EEIL and ICaRL. In conclusion, the ViT-TV approach offers robust support for the long-term intelligent development of the agricultural and food industry, especially showcasing significant applicability in continual learning for crop disease image recognition.

Enhanced melanoma detection using a fuzzy ensemble approach integrating hybrid optimization algorithm

The high death rate of melanoma makes it an important public health issue all over the world. Therefore, early detection of melanoma is crucial for a good prognosis. Deriving meaningful features from dermoscopic pictures, however, is difficult for a number of reasons, including a lack of training data, inconsistent classes, and intra-class variability. In order to solve this issue, we offer an automated technique for melanoma diagnosis from dermoscopic pictures by using high-level characteristics obtained from a powerful CNN architecture and LightGBM. The Ant Lion Optimization may get fixated on a local optimum rather than a global optimum as the problem's complexity rises. This may be a stumbling block to a more desirable ideal solution. The Hybrid Ant Lion Optimization (ALO) and Gray Wolf Optimization algorithm (GWO) combines elements from both ALO and GWO to benefit from their respective strengths and improve optimization performance. The HAM10000 dataset was utilized for this research, and it contains information on a variety of skin cancers, some of which are more common than others. The experimental analysis is done in python and the suggested model successfully distinguished between the two forms of skin cancer with an accuracy of 97.82%, recall of 96.80%, precision of 96.01%, and F1-score of 96.49%. The accuracy of ResNet is at 89.23, that of MLP at 90.87, of EfficientNet at 91.05, of ANN at 92.19, of IN3 at 93.73, and that of MobileNet at 94.90. The results shown that the suggested model outperforms baseline methods, providing substantial assistance to dermatologists and health specialists in the diagnosis of skin cancer.

Enhancing copy-move forgery detection through a novel CNN architecture and comprehensive dataset analysis

Enhancing copy-move forgery detection through a novel CNN architecture and comprehensive dataset analysis

Comparative analysis of transfer learning performance on generalised EEG data for use in a depression diagnosis task

The purpose of this work was to analyse the performance of different deep learning methods in the task of depression diagnosis based on bioelectrical brain activity data. In particular, to study the potential of transfer learning using an artificial neural network trained on a significant amount of “generalised” electroencephalography data in the task of diagnosing depression from non-invasive electroencephalogram signals. Methods. Deep learning approaches such as transfer learning and contrastive learning were used in the present study. Artificial neural networks were trained on the public HBN EO/EC task dataset containing recordings of electroencephalogram signals. The 1D CNN and EEGNet architectures were used as auxiliary artificial networks for transfer learning. In order to test the quality of contrastive learning, the dataset was augmented and the following algorithms were selected as the donor network: SimCLR, MoCo, NNCLR, BarlowTwins, DINO. Results. It was found that the EEGNet architecture used as a auxiliary network, due to its small size, does not give the full potential of contrastive learning algorithms. Therefore, EEGNet was replaced by a 1D CNN architecture with a larger number of parameters, which led to an increase in the quality performance of the models. Conclusion. Although the considered method of transient learning looks promising, the specificity of electroencephalogram signals and problems solved on their basis requires large-scale adaptation of algorithms and contrastive optimisation techniques for effective training of the target task. It is also worth noting the crucial role of the representativeness of the data set for training the donor network, since it is the completeness of real observations that increases the effectiveness of augmentation, which leads to an increase in the number of “useful” features in the latent space of the network and the best conditions for transfer learning in the target task. If we talk about the diagnosis of depression, the data should maximally represent examples of electroencephalograms of depressed patients.

An Ensemble Approach using Self-attention based MobileNetV2 for SAR classification

The Synthetic Aperture Radar (SAR) image classification has become an essential task in numerous applications. Despite its significance, SAR image classification remains challenging due to the inherent complexity and speckle noise characteristics of SAR data. To address this, an innovative approach is proposed in this study, combining the power of ensemble learning and the efficiency of MobileNetV2 architecture to achieve robust and accurate SAR image classification. The proposed deep learning model consists of two main components: an ensemble learning approach that combines multiple binary classifiers, and the employment of MobilenetV2 architecture with self-attention mechanism as a binary classifier. The ensemble approach uses a max-value selector for predicting final class from the individual classifier output. The MobileNetV2 architecture is finetuned by freezing weights and removing the last four layers. The inclusion of the attention layer within the architecture aids in refining the learning process by focusing on crucial features while disregarding irrelevant background elements in the images.. The effectiveness of the model is evaluated using Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark dataset with accuracy of classification as performance metric.The model is trained and tested with images captured at various aspect angles. This approach enhances the robustness of the results, making them more applicable to real-world scenarios when compared to prior methods. This study also compares the performance of the ensemble method using MobileNetV2 with state-of-art individual classifiers and different CNN architectures. The proposed ensemble method OVA-MbNN achieves an impressive classification accuracy of 76.63%, surpassing the performance of state-of-the-art classifiers while maintaining robustness.

Real-Time Convolutional Neural Networks for Emotion and Gender Classification

Emotion and gender recognition are important areas of research in the field of computer vision and human-computer interaction. The proposed CNN architecture is designed to extract features from facial images and classify them into six basic emotions (happy, sorrow, anger, fear, surprise, and disgust) and two genders (male and female) in real-time. To extract and categorize characteristics from facial photographs, the suggested CNN architecture consists of convolutional layers, pooling layers, and fully connected layers. The suggested system performs at the cutting edge for both emotion and gender recognition tasks when tested on publicly accessible datasets. The proposed real- time CNN architecture has potential applications in various fields, including social robotics, human-computer interaction, and affective computing.

Enhanced Accuracy of Sentiment Analysis of Facial Expressions Using Cnn Architecture

Enhanced Accuracy of Sentiment Analysis of Facial Expressions Using Cnn Architecture

Advanced Parkinson’s Disease Detection: A comprehensive artificial intelligence approach utilizing clinical assessment and neuroimaging samples

Medical experts are utilizing neuroimaging and clinical assessments to enhance the early identification of Parkinson’s disease. The current research initiative offers ways to identify Parkinson’s disease using machine learning and transfer learning. To carry out this, we extracted 7500 MRI images from 2022 and 2023 and 12 clinical assessment records from 2010 to 2023 from the well-known Parkinson’s Progression Marker Initiative (PPMI) database. Then, we applied machine and transfer learning approaches using clinical assessment records and MRI images, respectively. To identify Parkinson’s Disease (PD) using samples from clinical assessments, four distinct resampling techniques were employed. Subsequently, three machine learning models were applied to train on these resample records, and the recall score was analyzed. A hybrid of SMOTE and ENN proved to be the most effective approach for handling all of the imbalanced data, according to the recall study. Later, four different feature selection methods were used to find the top 10 features using these new samples. Lastly, we trained and validated the model using nine machine-learning algorithms. We also used explainable AI techniques like LIME and SHAP to interpret clinical assessment records. The extra tree classifier outperformed the others in terms of accuracy, reaching 98.44% using the tree-based feature selection technique. In addition to examining clinical assessment samples, this study investigated Parkinson’s disease using neuroimaging data. In pursuit of this objective, four pre-trained architectures were employed to analyze MRI images through two distinct approaches. The first approach involved utilizing the convolutional layer while replacing the remaining two layers with a customized Artificial Neural Network (ANN). Subsequently, training and evaluation are performed using our MRI samples, followed by analyzing significant weights using a LIME interpretable explainer. The second approach employs an improvisational technique without directly replacing the last layer. Instead, we predicted the weights of our MRI samples using the knowledge of the pre-trained model and stored them. Finally, CNN architecture was utilized for Parkinson’s disease detection, achieving an optimal accuracy of 85.08% with the implementation of DenseNet169 and CNN.

DEEPFAKE DYSTOPIA: NAVIGATING THE LANDSCAPE OF THREATS AND SAFEGUARDS IN MULTIMEDIA CONTENT

Deepfake technology, a fusion of deep learning and artificial intelligence, has emerged as a potent tool capable of crafting hyper-realistic yet entirely fabricated multimedia content. This comprehensive review explores the evolution, applications, and underlying principles of deepfake technology, emphasizing its potential implications for privacy, security, and the spread of misinformation. Using advanced deep learning algorithms, particularly Generative Adversarial Networks (GANs), deepfake technology manipulates facial features with remarkable precision, raising concerns about its malicious applications. The review examines the exponential growth in online content sharing facilitated by social media platforms and affordable devices, highlighting the convenience and accessibility but also the risks associated with the widespread use of deepfake technology. The core of deepfake technology, GANs, engages in an iterative competition between a discriminator and a generator, resulting in increasingly convincing synthetic data. A thorough review of the literature reveals significant research efforts focused on deepfake detection, leveraging techniques such as error-level analysis, CNN architectures, and hybrid approaches. The paper discusses regulatory measures, public awareness campaigns, and the critical role of digital forensic evaluation in mitigating deepfake threats. Challenges and concerns, including misinformation, privacy invasion, national security risks, and erosion of trust, are outlined. Mitigation strategies encompass advanced detection algorithms, regulatory frameworks, public awareness, and digital forensic evaluation, emphasizing the collaborative efforts required across technology developers, policymakers, the public, and digital forensic experts to navigate this evolving landscape and safeguard trust, privacy, and security in the digital age.

Rapid Prediction of Wastewater Index Using CNN Architecture and PLS Series Statistical Methods

Rapid Prediction of Wastewater Index Using CNN Architecture and PLS Series Statistical Methods

A Novel Architecture of Depthwise Separable CNN and Multi-Level Pooling for Detection and Classification of Myopic Maculopathy

A Novel Architecture of Depthwise Separable CNN and Multi-Level Pooling for Detection and Classification of Myopic Maculopathy