Variable-length deep convolutional neural networks by Internet Protocol Addresses Whale Optimization Algorithm for random and complex image classification

Abstract

While Deep Convolutional Neural Network (DCNN) is one of the most reliable deep learning methods for image classification, choosing the optimal DCNN architecture for a particular application can be quite challenging. This study aims to examine the usage of the Whale Optimization Algorithm (WOA) to discover the optimal structure for DCNNs automatically. Three advancements based on standard WOA are proposed to accomplish the objective. First, a novel Internet Protocol Address (IPA)-based coding approach is suggested, which makes it easier to encode DCNN layers utilizing whale vectors. The construction of variable-length DCNNs is then suggested using an Enfeebled layer to include certain whale vector dimensions. The learning procedure concludes by subdividing large datasets into smaller portions which are then arbitrarily assessed. In order to evaluate the proposed model’s effectiveness, it is compared to the performance of twenty-three different classifiers, including the state-of-the-art algorithm, on nine standard image classification benchmark datasets. The experimental results indicate that the suggested technique outperforms other benchmarks in 81 of 95 assessments. This variable-length approach is the first of its kind, utilizing WOA to evolve the topologies of DCNNs automatically.