Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (23): 154-164.DOI: 10.3778/j.issn.1002-8331.2206-0092

• Graphics and Image Processing • Previous Articles     Next Articles

Improved PSPNet Artillery Flame Segmentation Algorithm and Applications

ZHANG Wenwei, FU Qifan, WANG Yingguan, FU Hengcheng, WEI Zhi, DING Huaze   

  1. 1.Key Laboratory of Wireless Sensor Network and Communication,Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences, Shanghai 201800, China
    2.University of Chinese Academy of Sciences, Beijing 100864, China
  • Online:2023-12-01 Published:2023-12-01

改进PSPNet的炮弹火焰分割算法及应用

张雯玮,傅启凡,王营冠,傅衡成,魏  智,丁华泽   

  1. 1.中国科学院上海微系统与信息技术研究所 中国科学院无线传感网与通信重点实验室,上海 201800
    2.中国科学院大学,北京 100864

Abstract: For the current target range artillery flame image segmentation algorithm, there is a problem of poor segmentation effect on the flame boundary, which leads to a decrease in positioning accuracy. This paper proposes PSP_FPT(pyramid scene parsing feature pyramid with Transformer) algorithm based on PSPNet and combined with bi-directional feature fusion module and feature pyramid transformation module of full attention mechanism network. This algorithm can achieve high-precision segmentation of artillery fire targets. Firstly, the bi-directional feature fusion module is used to bi-directionally fuse the features after pyramid pooling. It enhances the spatial and semantic relevance of each sub-region as well as the global target to improve the accuracy of artillery fire segmentation. Secondly, in order to avoid the influence of smoke and dust around the flame on the segmentation effect, it designs the feature pyramid conversion module. It is combined with the full-attention mechanism network, which is able to optimize the feature mapping. To be more specific, the module can pay more attention to the region of interest in feature extraction, and enhance the spatial structure relationship between targets in the region simultaneously. At the same time, it improves the algorithm’s discriminative power between the artillery flame target and the background interference. In addition, the features after global pooling are used as the input of the full attention mechanism network, which solves the problem that there are a amount of parameters caused by full attention mechanism network since the long image input sequence. Additionally, it reduces the implementation difficulty of engineering applications. The experimental results show that the mIOU and mAcc of the algorithm achieves 98.01% and 98.97% respectively for segmentation on the artillery flame dataset, and has strong robustness and high accuracy for artillery flame segmentation.

Key words: artillery shell target identification and localization, high accuracy flame segmentation, bi-directional feature fusion, full attention mechanism, pyramid pooling feature

摘要: 针对目前靶场炮弹火焰图像分割算法对火焰边界分割效果差而导致定位精度下降的问题,基于PSPNet算法,结合双向特征融合模块以及全注意力机制网络的特征金字塔转换模块,提出改进PSPNet的炮弹火焰分割PSP_FPT(pyramid scene parsing_feature pyramid with Transformer)算法,实现对炮弹火焰目标的高精度分割。利用双向特征融合模块对金字塔池化后的特征进行双向融合,增强各子区域以及全局目标空间和语义特征的相关性,提升炮弹火焰分割的准确性。为了避免火焰周围烟雾、扬尘对分割效果的影响,将特征金字塔转换模块与全注意力机制网络相结合,优化双向特征融合模块输出后的特征映射,提升区域内目标之间的空间结构关系;提高算法对炮弹火焰目标与背景干扰之间的辨别力,进一步提高算法的识别能力。将全局池化后的特征作为全注意力机制网络的输入,解决了由于图像输入序列过长导致全注意力机制网络参数量过大的问题,进而降低工程应用的实现难度。实验结果表明,该算法在炮弹火焰数据集上分割的平均交并比达98.01%,平均准确率达98.97%,对炮弹火焰分割有较强的鲁棒性和较高的准确率。

关键词: 炮弹目标识别定位, 火焰高精度分割, 双向特征融合, 全注意力机制, 金字塔池化特征