A lightweight CNN architecture that achieves competitive performance without relying on pretrained models or transfer learning approaches, making it suitable for deployment on resource-constrained devices and balanced classification performance across all classes is developed.
Abstract
Image classification is an important area of computer vision and artificial intelligence that enables computers to automatically recognize and categorize visual information. This research aims to develop a Convolutional Neural Network (CNN)-based image classification model for recognizing six categories of natural and urban landscapes using the Intel Image Classification dataset from Kaggle. The preprocessing stage included image resizing, data augmentation, and pixel normalization to improve model generalization and reduce overfitting. The dataset was divided into 80% training data and 20% testing data. The proposed CNN architecture consists of four convolutional layers, max-pooling layers, and three fully connected dense layers with ReLU and Softmax activation functions. The novelty of this study lies in the development of a lightweight CNN architecture that achieves competitive performance without relying on pretrained models or transfer learning approaches, making it suitable for deployment on resource-constrained devices. Experimental results show that the model achieved 85.85% training accuracy and 85.47% testing accuracy. Performance evaluation using precision, recall, F1-score, and confusion matrix indicates balanced classification performance across all classes. Furthermore, the trained model was successfully converted into TensorFlow SavedModel, TensorFlow Lite, and TensorFlow.js formats to support cross-platform deployment. The findings demonstrate that the proposed CNN model is effective, efficient, and suitable for real-world landscape image classification applications.
The results indicate the successful implementation of the proposed AL-CNN model for reliable and accurate image classification, with results reported on a large-scale benchmark dataset that is widely accepted for performance evaluation.
M. Chawla, Rashmi Agrawal, Bharat Bhushan· Bulletin of Electrical Engin...· 0 citations
A comparative study of different deep learning architectures, including classical CNNs, deep hierarchical models, residual and dense networks, and compound-scaled architectures is presented, showing that deeper networks provide better representation, while residual connections and compound scaling improve training stability and efficiency.
Riyaz Mohammed· International Journal of App...· 0 citations
The development of artificial intelligence increases the risk of facial image manipulation through deepfake technology that is difficult to distinguish visually. This study aims to implement a Convolutional Neural Network (CNN) based on EfficientNetB0 to classify real and deepfake static facial images and integrate the model into a website application. The study follows the CRISP-DM framework using the FaceForencis++ extracted frames dataset from Kaggle, consisting of 60,000 images with 30,000 real images and 30,000 fake images. The data were divided into 70:15:15 ratios for training, validation, and testing, then processed through 224 x 224 pixel resizing, data augmentation, and facial area cropping using OpenCV Haar Cascade during inference. The model achieved 82.50% validation accuracy and 0.3874 validation loss at the 10th epoch. Decision threshold optimization at 0.78 produced 80.99% accuracy, 80.98% macro F1-score, and 78.56% recall for the fake class on the test data. The model was deployed in a Streamlit application, allowing users to upload facial images and receive detection results directly. The results indicate that the combination of EfficientNetB0 and threshold optimization can support preliminary verification of digital facial image authenticity.
Dimas Aditya Nugraha, Syariful Alam, Chandra Dewi Lestari· Jurnal Teknologi Informatika...· 0 citations
This study compares a custom CNN with three pre-trained models on the CIFAR-10 dataset, which comprises 60,000 colour images (32×32 pixels) across 10 categories, and shows that pre-trained models performed much better than the customised model.
Oluwadamilare (Asabia) Joseph Omoniyi, Omotosho Olawale Jacob, Ajaegbu Chigozirim et al.· International journal of res...· 0 citations
Because Generative AI, we now see highly realistic synthetic images that are difficult to distinguish from the real thing. This technology raises concerns about misinformation, the trustworthiness of digital content, and how people can misuse synthetic media. Therefore, finding an accurate and efficient way to detect AI-Generated images is a significant research challenge. In this research, we explore a hybrid image classification setup: we combine CNN-based feature extraction (using a pre-trained network) with SVM classification to determine which images are AI-Generated and which are real. For feature extraction, we use MobileNetV2, pre-trained on ImageNet, as the fixed extractor. Then, to classify these features, we use SVM with an RBF kernel. This experiment uses the Kaggle dataset AI Generated Images vs real Images, with a total of 975 images divided into training and testing sets with a stratified 80:20 split. Our model performs quite well: 84.1% accuracy, 81.8% precision, 91.7% recall, and an F1 score of 86.5%. The confusion matrix shows that the model is able to detect most of the AI-generated images, demonstrating its sensitivity to synthetic content. Overall, our CNN–SVM setup offers an effective and efficient alternative to full end-to-end deep learning methods for detecting these images. This holds promise for applications such as digital content verification, combating disinformation, and automated authenticity checking.
Hersatoto Listiyono, Yunus Anis, Sri Mulyani et al.· Engineering and Technology J...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
