Evolving Deep Convolutional Neural Networks by Extreme Learning Machine and Fuzzy Slime Mould Optimizer for Real-Time Sonar Image Recognition

Due to the shortcomings of conventional machine hearing methods in tackling data with high-dimension search space, such as the need for initial configuration and feature extraction manually, high complexity, and long processing time, real-time processing is challenging in the field of sonar image processing. Therefore, this paper proposes a hierarchical three-stage deep learning (DL)-based approach for real-time accurate sonar image recognition. A conventional deep convolutional neural network (DCNN) is used as an automatic feature extractor in the first stage. In the second stage, ELM replaces the last fully connected layer to reduce tuning and testing time. Due to the uncertainty imposed via the addition of ELM to the model, in the third stage, the slime mould algorithm (SMA) will be used to tune the input weights and biases of the ELM. Finally, fuzzy maps are used to balance the relationship between the SMA’s exploration and exploitation phases. For evaluating the efficiency of the designed fuzzy SMA (FSMA), we first use 23 standard benchmark mathematical optimization functions. Subsequently, we employ three experimental sonar datasets to examine the efficiency of DCNN–ELM–FSMA in dealing with high-dimensional datasets. For a comprehensive investigation, we compare FSMA to the standard SMA, whale optimization algorithm (WOA), gray wolf optimizer (GWO), kalman filter (KF), henry gas solubility optimization (HGSO), Harris Hawks optimization (HHO), chimp optimization algorithm (ChOA), genetic algorithm (GA), and particle swarm optimization (PSO), with respect to convergence rate, entrapment in local minima, and detection accuracy. The results demonstrate that the proposed strategy performs better in detecting underwater anomaly targets by an average of 2.11 percent compared to the best benchmark model.