Abstract:

In this paper, a one-stage anchor-free channel ship detection algorithm called FoveaSDet is proposed to address the problem of the small size ship, large scale transformation, and complex background in the image of drone aerial. FoveaSDet uses a backbone called SEResNeXt-I, which improves from ResNeXt, to increase the detection accuracy of small objects. Next, to solve the large scale transformation problem, the proposed method uses Foveahead to achieve anchor-free box object detection. The complete IOU loss has used to accomplish the bounding box regression so that the positioning accuracy of the detection box under a complex background can be raised. The experiment results show that the average precision and small objects’ average precision of the FoveaSDet is 71.6% and 47.0% testing on the real aerial image dataset. It has increased by 4.9% and 6.2% compared with FoveaBox, reflecting better detection accuracy and small object detection ability.

Key words: deep learning, object detection, feature extraction, anchor-free, loss function

摘要：

针对无人机航拍航道船舶影像中船舶目标较小、尺度变换大、背景复杂等问题，提出了一种基于FoveaBox网络的单阶段无锚框的航道船舶检测算法FoveaSDet。为提升小目标的检测精度，该算法使用基于残差网络改进的SEResNeXt-I作为骨干网。为改善尺度变换问题，FoveaSDet采用Foveahead实现无锚框目标检测。同时为提高复杂背景下检测框的定位精度，使用完全交并比损失实现边框回归。经实验测试，FoveaSDet算法在实景航拍数据集上的平均准确率（AP）和小目标准确率（APS）分别为71.6%和47.0%，相较于原始的FoveaBox提高了4.9%和6.2%，体现了更好的总体检测精度和小目标检测能力。

关键词: 深度学习, 目标检测, 特征提取, 无锚框, 损失函数