基于DenseNet和注意力机制的静爆场破片识别方法研究

Research on fragment identification of a static explosion field based on DenseNet and attention mechanism

针对静爆场破片着靶图像识别中对小破片的识别较为困难的问题,采用改进的SSD目标检测算法,将SSD网络模型中的骨干网络修改为DenseNet,减少网络参数,降低输入图像特征信息的消耗,最大程度地保留目标物体的细节信息。另外引入注意力机制模型,结合通道注意力机制和空间注意力机制,获取特征层通道和特征点的权值,充分提取小破片的特征信息。实验结果表明,提出的方法对小破片的检测准确率达到89.82%,与传统的SSD方法相比误检率提高了3.6%,漏检率提高了6%,为分析破片的飞散特性和毁伤效果提供了保障。

In view of the difficulty in the identification of small fragments in fragmenting target image recognition in a static blasting field, this paper uses an improved SSD target detection algorithm to modify the backbone network in the SSD model to DenseNet, which reduces the network parameters, reduces the consumption of the input image feature information, and preserves the details of the target objects to the greatest extent. In addition, the attention mechanism model is introduced, which combines channel attention mechanism and spatial attention mechanism to obtain the weights of the feature layer channels and feature points so as to fully extract the feature information of small fragments. The experimental results show that detection accuracy of the proposed method for small fragments reaches 89.82%. Compared with the traditional SSD method, the false detection rate increases by 3.6% and the missed detection rate increases by 6%, which provides a guarantee for analyzing the flying characteristics and damage effect of the fragments.