Intelligent fine-tuning of convolutional neural networks using flamingo search for traditional dance classification

ABSTRACT Developing efficient CNN architectures for image classification is computationally intensive, largely because identifying suitable hyperparameter combinations requires extensive experimentation. Traditional CNNs typically rely on fixed architectures and preset hyperparameters, reducing their ability to adapt efficiently when applied to diverse datasets. Previous studies have applied metaheuristic algorithms to tune CNN hyperparameters; however, many such methods optimize parameters independently or within rigid hierarchical schemes, limiting flexibility in overall network design. To address this issue, this research introduces a fine-tuned CNN architecture optimized through the Flamingo Search (FS) algorithm, a recent metaheuristic optimization method inspired by cooperative foraging behavior. The FS algorithm is employed to dynamically optimize the Neural Architecture Search Network (NASNet) CNN architecture, adjusting key hyperparameters such as the number of filters, activation functions, dropout rate, and optimizer settings. FS also explores alternative NASNet cell connections, enabling the model to adaptively construct an efficient architecture with improved feature extraction capability. The methodology involves two stages: (i) optimizing NASNET cell structures using FS on benchmark CIFAR-10 and CIFAR-100 datasets, and (ii) evaluating the optimized model on the Indian Classical Dance (ICD) dataset to test generalization performance. Implementation was carried out using Python and TensorFlow on an NVIDIA GTX 1060 GPU platform. Experimental results show that the proposed FS-CNN model achieves a classification accuracy of 97.5%, outperforming existing models such as HPSGW-CNN by 0.7%, while reducing computational cost to approximately 10 GPU days. These findings demonstrate that the FS-based optimization strategy provides an efficient and adaptive framework for large-scale image classification, with potential applicability to future tasks such as video classification and domain-specific recognition.