Fine-grained Classification Of Ship Images Based On Binary BCNN

In the military and maritime fields,using high-definition ship images provided by optical remote sensing satellites to quickly and accurately distinguish ships at a fine-grained level can play an important role in maritime traffic control,maritime rights protection,customs anti smuggling,maritime rescue,and homeland security maintenance.Deep neural networks have achieved good results in fine-grained image classification,but deep learning models typically require a large amount of computing resources and storage space.This limitation makes deploying and applying deep learning technology on resource limited devices such as optical remote sensing satellites a challenge.This study first uses the basic network structure of Res Net34 as the feature extractor of the bilinear convolutional neural network model,and adopts the high-order feature extraction ability of single source bilinear to improve the classification accuracy of the full precision network model as much as possible;Secondly,using binary neural networks in model quantization to compress the model and reduce the hardware requirements for network model deployment;Finally,in response to the issue of binary compression causing a large amount of information loss,starting from reducing quantization errors and adjusting network structure,scaling factors and group bilinear extraction methods were introduced to optimize the model.A fine-grained classification model for high-definition remote sensing ship images was proposed.The experiment shows that compared to the full precision network model,the binarized network model reduces its storage space by 3.8%;The optimized model has improved classification prediction accuracy by nearly 10 percentage points compared to before optimization.