Deformable Feature Pyramid Network for Ship Recognition

Ship recognition under complex sea environment and weather condition is a challenging task because the ship appearances change greatly, especially under large geometric transformation. The original feature pyramid network (FPN) [4] does not achieve good performance if it is implemented on ship detection directly, because it uses the fixed geometric structures in their building modules. In this paper, a deformable feature pyramid network is designed for ship recognition. The contributions are three folds: (1) We change the fixed geometric structure model of the original feature pyramid network to deformable geometric structure model and use the dilated convolution [12] instead of the original convolution. Correspondingly, deformable position-sensitive RoI pooling is used instead of the fixed geometric RoI pooling in the RoI-wise subnetwork. (2) The focal loss function [6] replaces the original mixed cross-entropy loss function. (3) Decay-NMS, a new post-processing method, is designed in this paper to improve the detection accuracy. The experimental results demonstrate the effectiveness and efficiency of our model.