Improved adaptive pooling techniques for convolutional neural networks in image classification

This paper presents the development and implementation of an Adaptive Convolutional Neural Network (ACNN) architecture that incorporates structured Mask-Pooling processing techniques to improve the adaptability and efficiency of pooling mechanisms using the CIFAR-10 dataset. The adaptive mask-pooling approach selects the pooling method based on the variance of the current 3x3 mask relative to the overall variance of the input image, thereby enhancing feature extraction and classification performance across various image datasets. Moreover, this paper improves the CNN model's performance and reduces training time by adjusting the stride value based on threshold conditions. The performance of the ACNN model is evaluated using a separate test dataset, focusing on key metrics such as precision, recall, and F1 score. To assess the model's accuracy, effectiveness, and robustness in image classification tasks, various techniques, including mean square error (MSE) and cross-validation, are employed. This paper enhances the understanding of feature representation and improves image classification by introducing novel adaptive mask pooling techniques within CNNs and comparing them to traditional pooling methods. The proposed ACNN model, which utilizes adaptive mask pooling, significantly boosts classification accuracy, robustness, and computational efficiency. These advancements offer various benefits in real-world applications, including medical imaging, autonomous vehicles, and safety systems. Extensive experiments conducted on collected datasets evaluate the effectiveness of the proposed methods, providing valuable insights into their potential impact on the implementation of CNNs in future image recognition applications. Furthermore, the results from adaptive mask clustering demonstrate superiority over traditional clustering methods, achieving maximum, average, precision, F1 score, recall, and MSE values of 0.70, 0.70, 0.75, 0.64, and 0.054, respectively. Additionally, the outcomes of adaptive mask clustering surpass those of traditional techniques by approximately 62%, 70%, and 89% for maximum, average, and frequency clustering, respectively.