Improved Convolutional Neural Networks for Image Classification Using Adaptive Dataset Splitting and Number of Filters Selection in Convolutional Layers

Résumé fourni par la source

Achieving high classification accuracy while maintaining low computational complexity remains a significant challenge for Convolutional Neural Network (CNN)-based models, particularly in time-critical image classification applications.This paper proposes an Adaptive Dual-CNN framework that improves classification performance through an adaptive dataset-splitting strategy.The proposed approach divides the input dataset into two sub-datasets based on image color characteristics using the Average Class Variance (ACV) and a dynamically computed Average Global Threshold (AGT).Images are then assigned to one of two specialized CNN models, enabling each model to learn features that are better suited to the characteristics of its corresponding sub-dataset.Unlike conventional single-CNN approaches, the proposed framework automatically computes the AGT from the input dataset and adaptively routes images to the most suitable CNN model without manual intervention.The Dual-CNN models are optimized by fine-tuning the number of convolutional filters, achieving an effective balance between classification accuracy and computational efficiency.The proposed framework was evaluated using the CIFAR-10 and Fashion-MNIST benchmark datasets.Experimental results demonstrate the effectiveness of the adaptive datasetsplitting strategy and the Dual-CNN architecture.The best-performing experiments achieved classification accuracies of 89% and 97.76% on CIFAR-10 and Fashion-MNIST, respectively.Compared with a conventional single-CNN model, the proposed framework improved classification accuracy by approximately 4-5% on CIFAR-10 and 3.76% on Fashion-MNIST.Furthermore, the proposed method achieved AUC values of 99.43% and 99.93% and mAP values of 95.49% and 99.28%, respectively.These results demonstrate that the proposed Adaptive Dual-CNN framework can significantly enhance classification performance while maintaining reasonable computational complexity.