Traditional Convolutional Neural Network Models Research Articles

Improvement The Accuracy of Convolutional Neural Network with Using Undersampling Method on Unbalanced Credit Card Dataset

In this study, we address the challenge of imbalanced data in credit card fraud detection by proposing a novel approach that leverages Convolutional Neural Networks (CNNs) and undersampling techniques. The imbalance in the dataset, typical of real-world financial transactions, often leads to biased models favoring the majority class. To mitigate this, we employ undersampling to balance the classes, thereby enhancing the CNN's ability to learn from minority instances crucial for fraud detection. Our method is validated on a large unbalanced credit card dataset, demonstrating significant improvements in accuracy compared to traditional CNN models trained on imbalanced data. We evaluate our approach using standard performance metrics, including precision, recall, and F1-score, showcasing its effectiveness in accurately identifying fraudulent transactions while minimizing false positives. Furthermore, we pro-vide insights into the CNN's decision-making process through visualization techniques, shedding light on its ability to discern fraudulent patterns within the data. Our findings highlight the importance of addressing class imbalance in fraud detection tasks and underscore the efficacy of undersampling in enhancing the performance of deep learning models, particularly CNNs, in handling imbalanced datasets.

A UAV Based Concrete Crack Detection and Segmentation Using 2-Stage Convolutional Network with Transfer Learning

This study explores a non-destructive testing (NDT) method for crack detection using a two-stage convolutional neural network (CNN) model, incorporating a combination of AlexNet and YOLO models through transfer learning. Crack detection is pivotal for assessing structural integrity and ensuring timely maintenance interventions. The developed model was rigorously tested in simulated environments and through physical experimentations with the use of a UAV to evaluate its effectiveness. A 2-stage model, based on AlexNet and YOLO, was developed for crack classification and segmentation. The developed model leveraged transfer learning to address limitations from traditional CNN models. A known dataset was used to evaluate the developed model, benchmarking it against other models. The classification network achieved an accuracy rate exceeding 90%, while the segmentation network successfully identified and delineated cracks in 85.71% of the images. Finally, the developed model was deployed using a UAV to perform crack detection and segmentation in a controlled environment. These results underscore the model's proficiency in both detecting and segmenting structural cracks, highlighting its potential as a reliable tool for enhancing the maintenance and safety of architectural structures. Doi: 10.28991/HIJ-2024-05-03-010 Full Text: PDF

Lightweight Network Bearing Intelligent Fault Diagnosis Based on VMD-FK-ShuffleNetV2

With the increasing complexity of mechanical equipment and diversification of deep learning models, vibration signals collected from such equipment are susceptible to noise interference. Moreover, traditional neural network models struggle to be effectively deployed in production environments with limited computational resources, severely impacting the accurate extraction and effective diagnosis of FK fault characteristics. In response to this challenge, this study proposes a fault diagnosis method for rolling bearings, integrating a lightweight ShuffleNetV2 network with variational mode decomposition (VMD) and the fast kurtogram (FK) algorithm. Initially, this paper introduces an enhanced FK method where the VMD algorithm is employed for data denoising, extracting FK post-denoising. These feature maps not only preserve critical signal information but also simplify data complexity. Subsequently, these feature maps are utilized to train and test the ShuffleNetV2 model, facilitating effective fault identification and classification. Ultimately, by conducting experimental comparisons with several mainstream lightweight network models, such as MobileNet and SqueezeNet, as well as traditional convolutional neural network models, this study validates the effectiveness of the proposed method in extracting fault characteristics from vibration signals, demonstrating superior diagnostic accuracy and computational efficiency. This provides a novel technical approach for health monitoring and fault diagnosis of industrial bearings and offers theoretical and experimental support for the deployment of lightweight networks in industrial applications.

High-noise solar panel defect identification method based on the improved EfficientNet-V2

As a crucial element in photovoltaic power generation systems, the condition of solar panels significantly impacts the efficiency of power generation. The ability to accurately and promptly detect defects in solar panels is essential for enhancing system performance. This study introduces a novel model for identifying defects in photovoltaic modules, leveraging an enhanced version of EfficientNet-V2. This model aims to address challenges in identifying defects in infrared images of solar panels under conditions of high-noise and low-model efficiency. To address the challenges of high image noise and blur, this article initially presents a methodology that combines the Db4 wavelet transform with a blind deconvolution algorithm for comprehensive preprocessing of the original image. Furthermore, this study optimizes the model's feature representation capabilities by implementing key transformations within the EfficientNet-V2 network framework. Notably, we replaced the traditional SE block with the more efficient channel attention (ECA) mechanism module. Due to its lightweight structure and effective performance, ECA substantially improves the model's capacity to extract complex and abstract image features, while also accelerating the training process's convergence speed and enhancing overall computational efficiency. At the classifier level, this paper innovatively integrates the XGBoost ensemble learning algorithm into the model, substituting the conventional softmax classifier used in traditional convolutional neural network (CNN). With its superior generalization capabilities, robust nonlinear modeling skills, and efficient computational characteristics, XGBoost can more accurately detect minute defects in solar panels based on the deep features produced by EfficientNet-V2, thereby significantly improving the accuracy and robustness of defect detection. Simulation results demonstrate that the proposed model structure outperforms traditional CNN models in terms of accuracy and stability, underscoring the efficacy of the enhanced EfficientNet-V2 model in detecting solar panel defects under high-noise conditions.

Enhancing the performance of convolutional neural network image-based steganalysis in spatial domain using Spatial Rich Model and 2D Gabor filters

Image-based steganalysis problem has attracted many researchers, and several solutions have been proposed. Deep learning-based methods are the most promising as they provide superior performance. Convolutional Neural network(CNN) based steganalysis methods are designed to improve the detection rate. Unlike traditional CNN models, CNN-based steganalysis requires careful design of preprocessing layers with filter initialization to obtain a good performance. In this paper, we established a CNN model that consists of two convolution layers for preprocessing and feature extraction, and four fully connected layers for classification. The preprocessing layer uses a set of efficient filter banks consisting of SRM and 2D Gabor filters. We conducted experiments using grayscale cover images from a popular and publicly available BOSSbase_1.01 database and Alask_v2 database with consideration for two different image sizes. The results showed that the proposed CNN model outperforms many state-of-the-art studies in two out of three well-known adaptive spatial domain steganography algorithms (S-UNIWARD, HUGO) and provides a close result for (WOW) algorithm when using the database with 512 × 512 images. On the other hand, the proposed model outperforms many state-of-the-art studies in the three algorithms when using the database with the original image size (256 × 256). Using image size 256, and the S-UNIWARD algorithm, the proposed model improved the detection accuracy rate by 13%, and 4.25% payloads of 0.2 and 0.4 bpp respectively compared to the previously best-known model (GBRAS-Net). The proposed model achieved 7.4% and 6.27% improvement in the detection accuracy for both payloads 0.2 and 0.4 bpp respectively using the HUGO algorithm compared with the previously best-known model (GBRAS-Net). For the WOW algorithm, the proposed model is slightly behind the best model (GBRAS-Net) but was able to obtain a close result for both payloads of 0.2 and 0.4 bpp, respectively. Using an image size of 512, the proposed model achieved 31.26%, 21.51%, 6.84%, 4.22%, and 1.96% improvement in the detection rate for the five payloads 0.1, 0.2, 0.3, 0.4, and 0.5 bpp respectively over S-UNIWARD algorithm compared to the previously best-known model (H-CNN). In addition, the proposed model achieved 27.60%, 23.69%, 12.66%, 5.27%, and 6.23% improved detection accuracy for the five payloads 0.1, 0.2, 0.3, 0.4, and 0.5 bpp respectively over HUGO algorithm compared with the previously best-known model (H-CNN). Finally, the proposed model provided 57.81%, 46.84%, 28.29%, 20.34%, and 13.79% improvement in the detection rate for the five payloads 0.1, 0.2, 0.3, 0.4, and 0.5 bpp respectively over WOW algorithm compared to the previously best-known model (H-CNN).

Remote sensing object detection with feature-associated convolutional neural networks

Neural networks have become integral to remote sensing data processing. Among neural networks, convolutional neural networks (CNNs) in deep learning offer numerous advanced algorithms for object detection in remote sensing imagery, which is pivotal in military and civilian contexts. CNNs excel in extracting features from training samples. However, traditional CNN models often lack specific signal assumptions tailored to remote sensing data at the feature level. In this paper, we propose a novel approach aimed at effectively representing and correlating information within CNNs for remote sensing object detection. We introduce object tokens and incorporate global information features in embedding layers, facilitating the comprehensive utilization of features across multiple hierarchical levels. Consideration of feature maps from images as two-dimensional signals, matrix image signal processing is employed to correlate features for diverse representations within the CNN framework. Moreover, hierarchical feature signals are effectively represented and associated during end-to-end network training. Experiments on various datasets demonstrate that the CNN model incorporating feature representation and association outperforms CNN models lacking these elements in object detection from remote sensing images. Additionally, integrating image signal processing enhances efficiency in end-to-end network training. Various signal processing approaches increase the process ability of the network, and the methodology could be transferred to other specific and well-defined task.

TRAFFIC SIGN RECOGNITION IN CHALLENGING WEATHER CONDITIONS USING CONVOLUTIONAL NEURAL NETWORKS

The ability of autonomous driving systems to recognize traffic signs in a variety of environmental conditions is crucial to their reliability. This study uses convolutional neural networks (CNNs) to provide a novel method for improving the accuracy of traffic sign recognition systems under challenging weather situations. The research focuses on developing a CNN-based model that is trained using the augmented German Traffic Sign Recognition Benchmark (GTSRB) dataset, which contains over 50,000 labeled images of road signs across 43 categories. To attempt to overcome the limitations of traditional CNN models under adverse environmental conditions, this work presents adaptive feature extraction layers that are especially made to reduce visibility problems brought on by rain, fog, and snow. By taking a comprehensive approach, the model uses advanced data augmentation methods to simulate different weather scenarios, greatly increasing the diversity of the training dataset. Through an analysis of theoretical and practical aspects, the study demonstrates how CNNs enhance the accuracy and efficiency of road sign detection systems in a different weather condition. Research evaluates the model's effectiveness using metrics such as precision, recall, and F1-score, proving its capability to reduce false positives and accurately detect relevant road sign instances. Additionally, the paper highlights the importance of careful dataset preparation and augmentation, model optimization, and training enhancements to improve the performance of road sign detection systems. The results of this research offer benefits for intelligent transport systems, autonomous driving, and road safety, pointing towards further advancements in precise and dependable road sign recognition technology.

Clivia biosensor: Soil moisture identification based on electrophysiology signals with deep learning

Research has shown that plants have the ability to detect environmental changes and generate electrical signals in response. These electrical signals can regulate the physiological state of plants and produce corresponding feedback. This suggests that plants have the potential to be used as biosensors for monitoring environmental information. However, there are current challenges in linking environmental information with plant electrical signals, especially in collecting and classifying the corresponding electrical signals under soil moisture gradients. This study documented the electrical signals of clivia under different soil moisture gradients and created a dataset for classifying electrical signals. Subsequently, we proposed a lightweight convolutional neural network (CNN) model (PlantNet) for classifying the electrical signal dataset. Compared to traditional CNN models, our model achieved optimal classification performance with the lowest computational resource consumption. The model achieved an accuracy of 99.26%, precision of 99.31%, recall of 92.26%, F1-score of 99.21%, with 0.17M parameters, a size of 7.17MB, and 14.66M FLOPs. Therefore, this research provides scientific evidence for the future development of plants as biosensors for detecting soil moisture, and offers insight into developing plants as biosensors for detecting signals such as ozone, PM2.5, Volatile Organic Compounds(VOCs), and more. These studies are expected to drive the development of environmental monitoring technology and provide new pathways for better understanding the interaction between plants and the environment.

Counterfeit Currency Detection Using Machine Learning

Counterfeit currency detection is a critical aspect of maintaining financial integrity. This paper introduces an innovative methodology for identifying counterfeit Indian rupee notes. By employing advanced image processing techniques and Convolutional Neural Networks (CNNs), specifically ResNet architecture, the proposed system achieves impressive accuracy in distinguishing genuine from counterfeit banknotes. The dataset, meticulously compiled through a combination of web scraping and synthetic printing, ensures robust training and evaluation. Experimental findings underscore the efficacy of the ResNet-based approach, highlighting its superiority over traditional CNN models. This research contributes significantly to bolstering currency security and trust in financial transactions, offering practical insights for combating counterfeit currency challenges.

A novel fish individual recognition method for precision farming based on knowledge distillation strategy and the range of the receptive field.

With the continuous development of green and high-quality aquaculture technology, the process of industrialized aquaculture has been promoted. Automation, intelligence, and precision have become the future development trend of the aquaculture industry. Fish individual recognition can further distinguish fish individuals based on the determination of fish categories, providing basic support for fish disease analysis, bait feeding, and precision aquaculture. However, the high similarity of fish individuals and the complexity of the underwater environment presents great challenges to fish individual recognition. To address these problems, we propose a novel fish individual recognition method for precision farming that rethinks the knowledge distillation strategy and the chunking method in the vision transformer. The method uses the traditional convolutional neural network model as the teacher model, introducing the teacher token to guide the student model to learn the fish texture features. We propose stride patch embedding to expand the range of the receptive field, thus enhancing the local continuity of the image, and self-attention-pruning to discard unimportant tokens and reduce the model computation. The experimental results on the DlouFish dataset show that the proposed method in this paper improves accuracy by 3.25% compared to ECA Resnet152, with an accuracy of 93.19%, and also outperforms other vision transformer models.

Enhanced Vision Transformer and Transfer Learning Approach to Improve Rice Disease Recognition

In the evolving landscape of agricultural technology, recognizing rice diseases through computational models is a critical challenge, predominantly addressed through Convolutional Neural Networks (CNN). However, the localized feature extraction of CNNs often falls short in complex scenarios, necessitating a shift towards models capable of global contextual understanding. Enter the Vision Transformer (ViT), a paradigm-shifting deep learning model that leverages a self-attention mechanism to transcend the limitations of CNNs by capturing image features in a comprehensive global context. This research embarks on an ambitious journey to refine and adapt the ViT Base(B) transfer learning model for the nuanced task of rice disease recognition. Through meticulous reconfiguration, layer augmentation, and hyperparameter tuning, the study tests the model's prowess across both balanced and imbalanced datasets, revealing its remarkable ability to outperform traditional CNN models, including VGG, MobileNet, and EfficientNet. The proposed ViT model not only achieved superior recall (0.9792), precision (0.9815), specificity (0.9938), f1-score (0.9791), and accuracy (0.9792) on challenging datasets but also established a new benchmark in rice disease recognition, underscoring its potential as a transformative tool in the agricultural domain. This work not only showcases the ViT model's superior performance and stability across diverse tasks and datasets but also illuminates its potential to revolutionize rice disease recognition, setting the stage for future explorations in agricultural AI applications.

Revolutionizing image recognition: The dominance and future potential of convolutional neural networks

As AI technology advances swiftly and diverse industries increasingly require image processing, traditional image recognition methods are displaying their limitations. This paper explores the evolving landscape of AI technology in the context of image processing and highlights the limitations of traditional image recognition methods. With the proliferation of big data and the evolution of deep learning, convolutional neural networks (CNNs) have emerged as a dominant solution for image recognition across diverse industries. The paper begins by elucidating the architecture of CNNs and introduces commonly employed traditional CNN models. Furthermore, it offers practical insights into the application of CNNs within various industries, illuminating the path for future CNN development. The transformative potential of CNNs is underscored, as they possess the ability to extract intricate patterns from images, reshaping numerous domains. The paper's primary focus is on CNNs in the realm of image recognition, encompassing efforts to enhance precision and efficiency in CNN-based image recognition, as well as addressing real-world challenges in this domain using CNNs.

Rolling bearing fault diagnosis based on Gramian angular difference field and improved channel attention model.

Fault diagnosis of rolling bearings is a critical task, and in previous research, convolutional neural networks (CNN) have been used to process vibration signals and perform fault diagnosis. However, traditional CNN models have certain limitations in terms of accuracy. To improve accuracy, we propose a method that combines the Gramian angular difference field (GADF) with residual networks (ResNet) and embeds frequency channel attention module (Fca) in the ResNet to diagnose rolling bearing fault. Firstly, we used GADF to convert the signals into RGB three-channel fault images during data preprocessing. Secondly, to further enhance the performance of the model, on the foundation of the ResNet we embedded the frequency channel attention module with discrete cosine transform (DCT) to form Fca, to effectively explores the channel information of fault images and identifies the corresponding fault characteristics. Finally, the experiment validated that the accuracy of the new model reaches 99.3% and the accuracy reaches 98.6% even under an unbalanced data set, which significantly improves the accuracy of fault diagnosis and the generalization of the model.

Efficient multimodel method based on transformers and CoAtNet for Alzheimer's diagnosis

Convolutional neural networks (CNNs) have been widely used in medical imaging applications, including brain diseases such as Alzheimer's disease (AD) classification based on neuroimaging data. Researchers extract the potential brain regions related to AD disease using CNN from various imaging modalities due to its architectural inductive bias. The major limitation of the current CNN-based model is that it doesn't capture long-range relationships and long-distance correlation within the image features. Vision transformers (ViT) have proven an astounding performance in encoding long-range relationships with strong modeling capacity and global feature extraction due to the self attention mechanism. However, ViT doesn't model the spatial information or the local features within the image and is hard to train. Researchers have demonstrated that combining CNN and a transformer yields outstanding results. In this study, two new methods are proposed for Alzheimer's disease diagnosis. The first method combines the Swin transformer with an enhanced EfficientNet with multi-head attention and a Depthwise Over-Parameterized Convolutional Layer (DO-Conv). The second method consists of modifying the CoAtNet network with ECA-Net and fused inverted residuals blocks. We evaluated the effectiveness of our proposed methods based on the Open Access Series of Imaging Studies (OASIS) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). Further, we evaluated the proposed methods using the Gradient-based Localization (Grad-CAM) method. The first method achieved 93.23% accuracy of classification on the OASIS dataset. The second method achieved 97.33% accuracy of classification on the OASIS dataset. We applied different multimodal image fusion methods (MRI and PET, MRI and CT) using our proposed method. The experimental results demonstrate that the fusion method based on PET and MRI outperforms the fusion method based on MRI and CT achieving 99.42% accuracy. Our methods outperform some traditional CNN models and the recent methods that are based on transformer for AD classification.

Efficient pollen grain classification using pre-trained Convolutional Neural Networks: a comprehensive study

Pollen identification is necessary for several subfields of geology, ecology, and evolutionary biology. However, the existing methods for pollen identification are laborious, time-consuming, and require highly skilled scientists. Therefore, there is a pressing need for an automated and accurate system for pollen identification, which can be beneficial for both basic research and applied issues such as identifying airborne allergens. In this study, we propose a deep learning (DL) approach to classify pollen grains in the Great Basin Desert, Nevada, USA. Our dataset consisted of 10,000 images of 40 pollen species. To mitigate the limitations imposed by the small volume of our training dataset, we conducted an in-depth comparative analysis of numerous pre-trained Convolutional Neural Network (CNN) architectures utilizing transfer learning methodologies. Simultaneously, we developed and incorporated an innovative CNN model, serving to augment our exploration and optimization of data modeling strategies. We applied different architectures of well-known pre-trained deep CNN models, including AlexNet, VGG-16, MobileNet-V2, ResNet (18, 34, and 50, 101), ResNeSt (50, 101), SE-ResNeXt, and Vision Transformer (ViT), to uncover the most promising modeling approach for the classification of pollen grains in the Great Basin. To evaluate the performance of the pre-trained deep CNN models, we measured accuracy, precision, F1-Score, and recall. Our results showed that the ResNeSt-110 model achieved the best performance, with an accuracy of 97.24%, precision of 97.89%, F1-Score of 96.86%, and recall of 97.13%. Our results also revealed that transfer learning models can deliver better and faster image classification results compared to traditional CNN models built from scratch. The proposed method can potentially benefit various fields that rely on efficient pollen identification. This study demonstrates that DL approaches can improve the accuracy and efficiency of pollen identification, and it provides a foundation for further research in the field.

A novel imbalanced dataset mitigation method and ECG classification model based on combined 1D_CBAM-autoencoder and lightweight CNN model

To address the problems of the imbalanced datasets, limited computing power, and memory capability of wearable devices, this paper proposes a new hybrid method for imbalanced dataset mitigation and a novel lightweight Convolutional Neural Network (CNN) framework for electrocardiogram(ECG) signal classification. First, data enhancement method is used by changing the amplitude of a small number of samples and combining with the weighted loss function in the deep learning model. Then, One-Dimensional Convolutional Block Attention Module (1D_CBAM) mechanism autoencoder model is proposed to compress the ECG heartbeat signals and extract the features of the ECG heartbeats. A modified one-dimensional lightweight CNN model based on MobileNet architecture is eventually used to categorize various arrhythmias utilizing the above features. Using the MIT-BIH arrhythmia dataset, the proposed model is tested by dividing the MIT Database(MITDB) into Dataset1 and Dataset2, in conjunction with suggestions from AAMI and other heartbeat signal classification methods. The experimental results show that the proposed imbalanced dataset mitigation method has higher performance than other methods in this paper, the proposed model has the most stable convergence and the lowest reconstruction loss (approximately 0.1) in the feature extraction process under the same conditions. During the classification process, the lightweight CNN model has approximately 4x fewer trainable parameters(Params: 12,788) and its performance [Dataset1: F1-Score: 96.24% ACC: 98.21%; Dataset2: F1-Score: 98.73% ACC: 99.61%] is comparable to the traditional CNN model [Dataset1: F1-Score: 96.98% ACC: 99.17%; Dataset2: F1-Score: 99.18% ACC: 99.72%].

Improved convolutional neural network chiller early fault diagnosis by gradient-based feature-level model interpretation and feature learning

For chillers, fault diagnosis (FD) is important for maintaining system reliability and performance. Deep learning methods, such as convolutional neural network (CNN), have been widely studied for chiller FD for its more significant diagnosis accuracy. But CNN model with deep layers and complex structures is black-box and difficult to interpret, which would greatly limit its practical FD applications for chillers. Traditional CNN model interpretation method may be not sensitive to interpret the chillers systematic faults especially at their early stages. Hence, to further obtain better interpretation of the CNN FD model, this study proposed a high-sensitivity gradient-based interpretation method. The proposed method adopts a softsign-forward-ReLU-backward manner to interpret the CNN model from the prospective of fault-discriminative feature, which localizes the fault-related feature variables and visualizing the diagnosis criteria for the CNN identified faults. The ASHRAE research project 1043 (RP-1043) chiller fault dataset was used to validate the proposed model interpretation and explanation method with higher feature-level sensitivity for the incipient fault. Based on the feature-level explanation and feature learning results, different feature combinations were investigated to improve diagnosis accuracy of some early-stage faults. If only small sizes of training data were available for modelling, 17 fault-related features were selected from the original 64 features to re-develop the CNN model and achieved diagnosis accuracy improvement of 9% for the early-stage improper refrigerant charge faults at most.

RamanCMP: A Raman spectral classification acceleration method based on lightweight model and model compression techniques

In recent years, Raman spectroscopy combined with deep learning techniques has been widely used in various fields such as medical, chemical, and geological. However, there is still room for optimization of deep learning techniques and model compression algorithms for processing Raman spectral data. To further optimize deep learning models applied to Raman spectroscopy, in this study time, accuracy, sensitivity, specificity and floating point operations numbers(FLOPs) are used as evaluation metrics to optimize the model, which is named RamanCompact(RamanCMP). The experimental data used in this research are selected from the RRUFF public dataset, which consists of 723 Raman spectroscopy data samples from 10 different mineral categories. In this paper, 1D-EfficientNet adapted to the spectral data as well as 1D-DRSN are proposed to improve the model classification accuracy. To achieve better classification accuracy while optimizing the time parameters, three model compression methods are designed: knowledge distillation using 1D-EfficientNet model as a teacher model to train convolutional neural networks(CNN), proposing a channel conversion method to optimize 1D-DRSN model, and using 1D-DRSN model as a feature extractor in combination with linear discriminant analysis(LDA) model for classification. Compared with the traditional LDA and CNN models, the accuracy of 1D-EfficientNet and 1D-DRSN is improved by more than 20%. The time of the distilled model is reduced by 9680.9s compared with the teacher model 1D-EfficientNet under the condition of losing 2.07% accuracy. The accuracy of the distilled model is improved by 20% compared to the CNN student model while keeping inference efficiency constant. The 1D-DRSN optimized with channel conversion method saves 60% inference time of the original 1D-DRSN model. Feature extraction reduces the inference time of 1D-DRSN model by 93% with 94.48% accuracy. This study innovatively combines lightweight models and model compression algorithms to improve the classification speed of deep learning models in the field of Raman spectroscopy, forming a complete set of analysis methods and laying the foundation for future research.

Images Classification Integrating Transformer with Convolutional Neural Network

Convolutional neural networks (CNN) are one of the most widely used deep learning methods in computer vision, which can effectively extract local spatial information from images, but lack global understanding and dependency modelling of image features. As a result, contextual information cannot be fully utilized by the network. For example, on coordinate modelling tasks (such as object detection, image generation, etc.), CNN may not be able to accurately locate or reconstruct the position and shape of objects. In contrast to traditional CNN models such as ResNet, Transformers rely on their global attention mechanism to capture long-distance dependencies between patches. The thesis presents an enhanced lightweight method which integrates Transformer with five convolutional neural layers. Model based on CNN and Transformer is tested on the two benchmark datasets MNIST and CIFAR-10. After a few epochs, the model is convergent and reaches high accuracy of 99.34% in MNIST and 92.04% in CIFAR-10. This model outperforms the single CNN and some state-of-the-art models in classifying both datasets, especially in distinguishing similar images like ‘6’ and ‘9’, ‘bird’ and ‘plane’. These results indicate the model's good robustness and generality.

Two-step hierarchical neural network for classification of dry age-related macular degeneration using optical coherence tomography images.

The aim of this study is to apply deep learning techniques for the development and validation of a system that categorizes various phases of dry age-related macular degeneration (AMD), including nascent geographic atrophy (nGA), through the analysis of optical coherence tomography (OCT) images. A total of 3,401 OCT macular images obtained from 338 patients admitted to Shenyang Aier Eye Hospital in 2019-2021 were collected for the development of the classification model. We adopted a convolutional neural network (CNN) model and introduced hierarchical structure along with image enhancement techniques to train a two-step CNN model to detect and classify normal and three phases of dry AMD: atrophy-associated drusen regression, nGA, and geographic atrophy (GA). Five-fold cross-validation was used to evaluate the performance of the multi-label classification model. Experimental results obtained from five-fold cross-validation with different dry AMD classification models show that the proposed two-step hierarchical model with image enhancement achieves the best classification performance, with a f1-score of 91.32% and a kappa coefficients of 96.09% compared to the state-of-the-art models. The results obtained from the ablation study demonstrate that the proposed method not only improves accuracy across all categories in comparison to a traditional flat CNN model, but also substantially enhances the classification performance of nGA, with an improvement from 66.79 to 81.65%. This study introduces a novel two-step hierarchical deep learning approach in categorizing dry AMD progression phases, and demonstrates its efficacy. The high classification performance suggests its potential for guiding individualized treatment plans for patients with macular degeneration.