融合多尺度特征与软注意力的态势认知方法

A Method of Situation Awareness Combining Multi-scale Features andSoft Attention

针对战场态势变化快、体量大导致指挥员难以快速认知这一问题,提出了一种融合多尺度特征与软注意力机制(MSF-SAM)的战场态势认知方法,其中,特征提取网络通过对不同卷积层的输出进行操作,并结合软注意力机制分配特征的不同权重,从而获得网络对图像不同部分的关注度,即上下文信息,并将上下文信息输入到LSTM网络中进行解析,最后得到图像态势认知结果。在兵棋态势图像数据集上进行了验证,通过比较该方法与经典方法、有无软注意力的对比方式,验证了所提方法的有效性,且在智能兵棋推演中具有一定的应用价值。

Aiming at the problem of rapid changes and the large size in the battlefield situation that make it difficult for commanders to quickly recognize the situation,a battlefield situation recognition method(MSF-SAM)that combines multi-scale features and soft attention mechanism is proposed,in which the feature extraction network operates on the output of different convolutional layers,and combines the soft attention mechanism to assign different weights of features to obtain the network’s attention to different parts of the image,context information,and to input the context information into the LSTM network for analysis,and finally gets the image situation recognition results.In addition,the method is verified on the wargame situation image datasets.The proposed method in this paper and the classic method are compared,and soft attention and non-attention mode are also compared,the effectiveness of the method is verified,and it has some application value in intelligent wargaming.