Preliminary results of applying neural networks to ship image recognition

Abstract

Summary form only given, as follows. A set of 39 pictures of four ship models in various positions was collected. The pictures were preprocessed to remove position and scale variations. In each picture (40*100 pixels) the ship image extended to both sides of the picture or from top to bottom. A subset of these pictures was used to train a large neural network (NN) using the generalized delta rule learning algorithm. The NN was tested on both the original images and simulated mirror images of the ships. When the maximum output from both presentations was used for making a classification decision, the NN successfully recognized the ships in all positions. It is observed that using first-layer weights initialized to zero produces faster learning and better performance than networks using only randomized weights.<>