IMAGE-BASED MALWARE CLASSIFICATION WITH COMPACT CONVOLUTIONAL TRANSFORMER

Plant disease detection is a critical phenomenon, which may enable farmers to reduce losses by increasing productivity. The ability of Convolutional Neural Networks (CNNs) is still limited when dealing with small datasets, which impacts their broader application. A model is prone to overfitting when dealing with a few samples. This limits generalization and reduces its performance since the model may be memorizing data instead of learning general patterns. We used the regularization technique known as early stopping to avoid overfitting and obtain a better generalization. Early stopping ensures the model to generalize and perform well on new data. The EfficientNetB0 architecture — one of the highly advanced CNN architectures — to achieve greater accuracy with fewer parameters. Coupled with the early stopping technique integrated into our approach, to achieve improved generalization for small and diverse datasets.

This study systematically evaluates and compares the effectiveness of two deep learning architectures, VGG16 and ResNet50, in automating electrocardiogram (ECG) image classification for cardiac condition diagnosis. The dataset was obtained from a public source and consists of 2,898 color ECG images converted from raw signals, categorized into four classes: Abnormal Heartbeat, Myocardial Infarction, Normal Individual, and History of Heart Attack. The data underwent preprocessing steps including resizing to 224×224 pixels, pixel normalization to a 0–1 range, label encoding, one-hot encoding, and an 80:20 split for training and testing. Transfer learning was applied using feature representations from the VGG16 and ResNet50 models, employing the Adam optimizer and categorical cross-entropy loss function. To enhance training efficiency and prevent overfitting, early stopping was implemented based on validation loss performance. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. The results showed that VGG16 achieved 95% accuracy with a loss of 0.1522, precision of 95%, recall of 94%, and F1-score of 94%. In contrast, ResNet50 attained 81% accuracy with a loss of 0.5730, precision of 82%, recall of 79%, and F1-score of 80%. These findings indicate that, within the context of this study, VGG16 consistently outperformed ResNet50 across all evaluation metrics in the ECG image classification task. Therefore, the application of transfer learning using the VGG16 model demonstrates strong potential as an effective approach for AI-based ECG image classification systems.

Deep neural-kernel blocks

The BERT pre-trained language model has achieved good results in various subtasks of natural language processing, but its performance in generating Chinese summaries is not ideal. The most intuitive reason is that the BERT model is based on character-level composition, while the Chinese language is mostly in the form of phrases. Directly fine-tuning the BERT model cannot achieve the expected effect. This paper proposes a novel summary generation model with BERT augmented by the pooling layer. In our model, we perform an average pooling operation on token embedding to improve the model’s ability to capture phrase-level semantic information. We use LCSTS and NLPCC2017 to verify our proposed method. Experimental data shows that the average pooling model’s introduction can effectively improve the generated summary quality. Furthermore, different data needs to be set with varying pooling kernel sizes to achieve the best results through comparative analysis. In addition, our proposed method has strong generalizability. It can be applied not only to the task of generating summaries, but also to other natural language processing tasks.

Road markings play an important role in vehicle navigation, helping as visual signs that promote road safety and compliance with traffic rules. This study uses quantitative experimental method to compare a baseline Convolutional Neural Network (CNN) trained from scratch with transfer learning using EfficientNetB0 using a small custom dataset captured via purposive sampling. The dataset was divided into training (70%), validation (30%), and a separate test set (75 images) with balanced class distribution. Results show that the baseline CNN achieved 56.3% validation accuracy and 66.7% test accuracy (macro F1-score: 0.672), while Effi- cientNetB0 reached 100% validation accuracy and 93.3% test accuracy (macro F1-score: 0.932). The 26.6 percentage point improvement in test accuracy highlights the advantage of transfer learning on small datasets, extending previous work by providing multi-class classification of pedestrian crossings, speed bumps, and lane dividers—a task that had not previously been considered in a single system.

Deploying convolutional neural networks in resource-constrained settings often requires sacrificing accuracy for efficiency, particularly under small-data regimes. MobileNetV2 is a leading lightweight backbone, but its performance on fine-grained classification tasks can suffer when training data is scarce and compute budgets are limited. This study proposes a unified training optimisation pipeline for MobileNetV2 that integrates five complementary techniques: transfer learning, dynamic data augmentation, dropout regularisation, adaptive learning-rate scheduling, and early stopping. This approach was evaluated on a subset of the Stanford Dogs dataset (1290 images, eight breeds) and compared against a standard MobileNetV2 baseline as well as MobileNetV3Small and EfficientNetB0 under identical training conditions. Performance is assessed via Top-1 accuracy, training/inference time, and statistical significance (ANOVA, Tukey’s HSD, Cohen’s d), with an ablation study isolating the contribution of each component. The optimised pipeline reduces training time by 50–70% (609 s vs. 1213 s) and inference latency by 10% (0.028 s vs. 0.031 s per image), while incurring only a 1.4% drop in Top-1 accuracy (98.6% vs. 100.0%). Statistical tests confirm significant efficiency gains (p < 0.001) and marginal accuracy loss (p = 0.0008). Ablation analysis demonstrates that no single technique achieves these combined benefits. Under identical conditions, MobileNetV2 outperforms alternative lightweight models that fail to converge on this small dataset. Coordinating established optimisation strategies yields a robust, resource-efficient classifier well suited to edge-computing and on-device vision tasks with limited data. This pipeline offers a practical blueprint for deploying deep learning models in small-data, low-power environments.

Question classification is a primary essential study for automatic question answering implementations. Linguistic features take a significant role to develop an accurate question classifier. Recently, deep learning systems have achieved remarkable success in various text-mining problems such as sentiment analysis, document classification, spam filtering, document summarization, and web mining. In this study, we explain our study on investigating some deep learning architectures for a question classification task in a highly inflectional language Turkish that is an agglutinative language where word structure is produced by adding suffixes (morphemes) to root word. As a non-Indo-European language, languages like Turkish have some unique features, which make it challenging for natural language processing. For instance, Turkish has no grammatical gender and noun classes. In this study, user questions in Turkish are used to train and test the deep learning architectures. In addition to this, the details of the deep learning architectures are compared in terms of test and 10-cross fold validation accuracy. We use two major deep learning models in our paper: long short-term memory (LSTM), Convolutional Neural Networks (CNN), and we also implemented the combination of CNN-LSTM, CNN-SVM structures and a number of various those architectures by changing vector sizes and the embedding types. As well as this, we have built word embeddings using the Word2vec method with a CBOW and skip gram models with different vector sizes on a large corpus composed of user questions. Our another investigation is the effect of using different Word2vec pre-trained word embeddings on these deep learning architectures. Experiment results show that the use of different Word2vec models has a significant impact on the accuracy rate on different deep learning models. Additionally, there is no Turkish question dataset labeled and so another contribution in this study is that we introduce new Turkish question dataset which is translated from UIUC English question dataset. By using these techniques, we have reached an accuracy of 94% on the question dataset.

Maize is an essential cereal for humans and animals worldwide, and it is one of the staple food in Kenya. One of the main challenges facing the maize crop in Kenya is the presence of diseases spreading quickly. Early recognition of maize pathogen and disease help at preventing the disease from spreading throughout the field. This paper proposes a regularized Multitask learning (MTL)–Convolutional Neural Networks (CNN) model for simultaneously identifying maize disease and its pathogen from diseased maize images. MTL allows training one model for multiple tasks at a time, which may improve the accuracy of each task by taking advantage of their commonalities. Our baseline is made of two CNN classification models, one of them being overfitting. We then build an MTL based on the two models, which increases the test accuracy of the overfitting model from 60.08% to 74.48%. The results show that the accuracy rises to 77.44% while combining MTL to the Early stopping method. However, the test accuracy goes up to 85.22 percent when MTL is combined with Early Stopping and Transfer Learning. The model is deployed to an android mobile application for maize farmers as end-users which is very important for costs reduction and time saving.

A MACHINE-LEARNING APPROACH FOR PREDICTING CHEST WALL ACCELERATION FROM WEARABLE CARDIOVASCULAR INSTRUMENT SIGNALS

This paper proposes an efficient and cost effective deep learning architecture to detect the diabetic retinopathy in real time. Diabetes is a leading root cause of eye disease in patients. It illuminates eye vessels, and releases blood form vessels. Early detection of diabetic retinopathy is useful to reduce the risk of blindness or any hazard. In this paper, after some preprocessing and data augmentation, Inception V3 is used as pre-trained model to extract the initial features set. Convolutional neural network has been used with attention layers. These additional CNN layers are added to extract the deep features to improve classification performance and accuracy. Initially, the model was proposed by Kevin Mader in Kaggle. The paper introduced additional layers in proposed model and improved the validation and testing accuracy significantly. More than 90% validation accuracy was achieved with the proposed Convolutional Neural Network model. Testing accuracy was improved up to 5%. This improvement in accuracy is very significant because the dataset is imbalanced and contains noisy images. It is concluded that global average pooling (GAP) based attention mechanism increased deep learning architecture accuracy to detect the Diabetic Retinopathy in imbalanced and noisy image dataset

Pooling layers are essential in convolutional neural networks (CNNs) for reducing data size while preserving key features. Traditional methods such as Max and Average pooling have limitations. Max pooling is sensitive to noise, while Average pooling treats all activations equally. Although T-Max-Avg pooling addresses these limitations through adaptive top-k selection, its rigid decision rule requires multiple threshold comparisons and limits efficiency, motivating a simpler decision mechanism. This study introduces average-thresholding pooling (ATP), a simplified adaptive method that replaces multiple threshold comparisons with a single decision based on the average of the top-k activations. This design improves computational efficiency and reduces sensitivity to outliers. Experiments on the STL-10 dataset using a LeNet-5 architecture show that the proposed method achieves accuracy comparable to T-Max-Avg pooling (~55.5%) while consistently improving both training efficiency and inference speed. These results indicate that ATP provides a lightweight and practical alternative for CNN-based image classification, offering an improved balance between classification performance and computational efficiency.

Neighbour feature attention-based pooling

Syntactic methods in computer vision represent visual patterns in a hierarchical and compositional perspective, which is converted to strings. Long short-term memory (LSTM) is able to learn patterns in sequences. In this letter, we propose a syntactic approach to represent visual patterns as sequences of symbols, and we use an LSTM as a classifier to learn the relationship between the symbols in sequences. An extensive experimental evaluation using aerial images from a soybean field captured by unmanned aerial vehicles has been conducted to compare our method with two deep learning architectures, one syntactic method, and one shallow learning algorithm. The results achieved by the proposed method maintain stability even when trained on small data sets, suggesting that representing visual patterns in a compositional way, repeating primitives, may be a viable alternative when there are only a limited number of samples for training.

An essential noninvasive medical diagnostic technique is magnetic resonance imaging (MRI), which is particularly useful for identifying brain cancers. While earlier algorithms proved effective on smaller MRI datasets, their performance suffered on bigger datasets. This study addresses the need for a swift and reliable brain tumor classification system capable of sustaining optimal performance across comprehensive MRI datasets. The convolutional neural network is implemented using the Keras library, incorporating the ResNet50 architecture as a pre-trained model. The ResNet50 model is fine-tuned for the specific brain tumor classification task, with a Global Average Pooling layer, dropout, and a final dense layer with softmax activation. Data augmentation techniques are employed to enhance the model’s robustness, including rotation, width and height shifts, and horizontal flips. The training process involves optimizing the model using the Adam optimizer with a learning rate of 0.0001. Early stopping, learning rate reduction on plateau, and model checkpointing are implemented as callbacks to ensure efficient training and prevent overfitting. The proposed model achieves a remarkable accuracy of 99.28 percent after 15 epochs. The classification task involves distinguishing among four classes: glioma, meningioma, pituitary, and no tumor.

Convolutional neural networks (CNN) is widely used for traffic sign recognition. Meanwhile, the compressive sensing technology is developing and applied to the field of image reconstruction in the compressive sensing domain. Therefore, we first propose a traffic sign recognition algorithm based on compressive sensing domain and convolution neural networks for traffic sign recognition. The algorithm converts the image into a compressed sensing domain through the measurement matrix without reconstruction, and can extract the discriminant nonlinear features directly from the compressed sensing domain. In order to improve the accuracy of traffic sign recognition, we further propose a cross-connected convolution neural networks (CCNN). Cross-connected convolution neural networks is a 9 layers framework with an input layer, six hidden layers (i.e., three convolution layers alternating with three pooling layers), a fully-connected layer and an output layer, where the second pooling layer is allowed to connect directly to the fully-connected layer across two layers. Experimental results on well-known dataset show that the algorithm improves the accuracy of traffic sign recognition. The recognition of our algorithm is even possible at low compressive sensing measurement rates.