A Novel Enhanced EfficientNet Model for Identification of Floating Debris on Marine Surface

Two methods of encoding time series, the Gramian Angular Summation Field (GASF) and the Gramian Angular Difference Field (GADF), are investigated in this research to provide a novel framework for encoding radar echo Doppler spectral patterns into pictures. The purpose of this study is to use the advantages offered by CNN models in order to better identify submerged objects. In order to improve the model's convergence, this study implements the coordinate attention (CA) method into EfficientNet's mobile inverted bottleneck convolution (MBConv) structure. The goal is to draw attention to the crucial role played by the coordinates in which the GAF map stores texture data. At last, the updated EfficientNet model is put through its paces by being tested and trained on the newly built GADF and GASF datasets, respectively. The findings from the experiments provide evidence that the suggested algorithm is successful. The upgraded EfficientNet model achieves a recognition accuracy of 96.13% on the GADF dataset and 96.28% on the GASF dataset, respectively. This is an increase of 1.74% and 2.06% when compared to the recognition accuracy of the pre-enhanced network model. The suggested approach achieves a greater level of accuracy than the traditional method of picture categorization while maintaining a smaller burden of computing.