Asymmetric Convolution-Based Neural Network for SAR Ship Detection from Scratch
Long Han, Da Ran, Wei Ye, Xu Dong Wu · 2020
In recent years, an increasing number of synthetic aperture radar (SAR) researchers have applied convolutional neural network (CNN)-based optical image object detection methods to the field of SAR ship detection via transfer learning technology, and achieve well performance. However, such approaches need to load a pre-trained model for initialization, which results in the structure of CNN is fixed and it is hardly for improvement and optimization; besides, there is domain mismatch for SAR ship detection, which restricts the detection performance to some extent. In the paper, we designed two asymmetric convolution blocks that can be easily embedded in any object detection methods, namely asymmetric and square convolution feature aggregation block (A-S AB) and asymmetric and square convolution feature fusion block (A-S FB), and embedded them in the classic DSOD by replacing all of the 3 × 3 convolution layers with A-S AB and A-S FB, respectively. Experiments on the RDISD_SAR dataset demonstrate that the A-S AB contribute up to 2.86% gain of average precision (AP) while significantly reducing the number of parameter and the amount of computation; the A-S FB contribute about 1.00% AP gain at channel decay factors of 0.2, 0.3, 0.5, and 0.6, and the number of parameter and the amount of computation are also reduced to varying degrees. Compared with the original DSOD, the well performance of the two structures designed in this paper is obvious.