Deep Convolutional Neural Network Architectures for Tonal Frequency Identification in a Lofargram

Advances in convolutional neural networks (CNNs) have driven the development of computer vision. Recent CNN architectures, such as those with skip residual connections (ResNets) or densely connected architectures (DenseNets), have facilitated backpropagation and improved the performance of feature extraction and classification. Detecting objects in underwater environments by analyzing sound navigation and ranging (sonar) signals is considered an important process that should be automated. Several previous approaches have addressed this challenge; however, there has been no in-depth study of CNN architectures that effectively analyze sonar grams. In this paper, we have presented the identification of tonal frequencies in lofargrams using recent CNN architectures. Our study includes 175 CNN models that are derived from five different CNN architectures and 35 different input patch sizes. The study results showed that the accuracy of the best model was as high as 96.2% for precision and 99.5% for recall, with an inference time of 0.184 s.