Novel Capsule Network Algorithm Based Image Classification with Efficiency

Despite the significant impact of convolutional neural networks (CNNs) across various domains, one of their limitations is the lack of consideration for spatial orientations. Capsule networks have the potential to address these issues by incorporating the storage and routing of pose information for the features they extract throughout their architectures. This involves seeking agreement between the lower-level feature extraction and the higher-level poses at each layer. Through the storage and routing of posture information across their architectures, capsule networks aim to resolve these issues. The fundamental components of the trainable segmentation engine include an encoder network, a corresponding decoder network, and a pixel-wise classification layer. The quantitative evaluations conducted demonstrate that Capsule Networks exhibit robust performance, competitive inference times, and superior memory efficiency compared to alternative architectures. The training and testing process utilizes datasets such as MNIST and CIAFR-10 datasets. The obtained results indicate that the proposed approach successfully outperforms CNNs in addressing image classification problems.