Deep neural network with transfer learning in remote object detection from drone

In this article we present a model of new deep learning composition for remote ship detection. Proposed architecture is composed of newly developed derivatives of ResNet, DenseNet and CNN composed into one global classifier. Since training of such model is demanding we have developed also a new proposition of transfer learning. Each of architectures was trained on different input data. In the final phase they are all composed into one global model for which training is finished by the use of augmented images from all the input collections. The proposed model of training enabled improved features of classification. Results of numerical experiments have shown that our newly proposed deep learning classifier with developed transfer learning model presents values of 99% Accuracy, 98% Precision, 99% Recall and 97% Specificity after training in only 40 iterations.