基于注意力机制的SOLOA船舶实例分割算法

SOLOA ship instance segmentation algorithm based on attention

目前,可见光船舶图像的实例分割仍然是较有挑战性的工作。由于船舶图像场景复杂多变,多数实例分割算法无法对复杂场景下的船舶图像进行有效分割。本文提出了基于注意力机制的依靠位置分割目标(attention based segmenting objects by locations,SOLOA)船舶实例分割算法,针对网络特征图里实例信息不完善、背景干扰较多的问题,使用空间注意力机制来充分利用分类特征中的实例信息,建模图像实例间的相互关系并与分割特征相融合。实验结果表明,在新构建的船舶图像数据集上进行训练和测试,改进的网络模型能有效地增强网络特征中的实例信息、减弱背景的干扰。SOLOA算法的船舶实例分割准确率高于其他算法,可以很好地适应复杂场景下的船舶分割需求。

At present,the instance segmentation of visible ship images remains a highly challenging task.Most instance segmentation algorithms cannot effectively segment ship images in complex scenes due to the intricate and variable nature of ship images.A segmenting objects by locations based on attention(SOLOA)algorithm for ship instance segmentation,which utilizes the spatial attention mechanism to maximize the instance information in the classification features,is proposed in this paper.Here,the interrelationships between the image instances are modeled and fused with segmentation features.Training and testing results of the newly constructed ship image dataset show that the improved network model can effectively enhance the instance information in the network features and reduce the background interferences.The accuracy of ship instance segmentation by the SOLOA algorithm is higher than that of other algorithms;hence,the proposed algorithm can be effectively adapted to meet the demands of ship segmentation in complex scenes.