摘要
战场野外复杂红外场景中,由于背景灰度分布无规律、目标边缘模糊且纹理特征缺失,目标极易混淆在背景之中;由于嵌入式平台算力的限制,多数深度学习类检测算法难以应用于便携设备,无法实现快速有效的目标识别。提出一种基于运动目标提取与高效机器学习模型结合的目标识别方法:通过运动检测实现目标像素级分割,经形态学处理后,定位单体目标;根据嵌入式平台算力高低,选择轻量化深度网络特征或轮廓特征,训练softmax模型,实现目标分类识别。将算法移植于嵌入式平台,对开源红外图像序列进行目标识别实验,实现多目标同时定位与分类,处理速度达56FPS。实验结果表明,该方法可对复杂背景中的红外目标进行实时有效识别。
In the complex infrared scene of the battlefield, the target is easily confused in the background due to the irregular gray level distribution, the blurred edge of the object and the lack of texture features. Due to the limitation of the computing performance of the embedded platform, most of the deep learning detection algorithms are difficult to apply to portable devices and cannot achieve fast and effective object recognition. An object recognition method based on moving object extraction and efficient machine learning model is proposed. This method firstly realizes the target pixel-level segmentation through motion detection, and locate the single target after morphological processing. Then select lightweight deep network features or contour features according to the computing performance of the embedded platform, train softmax model to achieve object classification. This algorithm is transplanted to the embedded platform, and the object recognition experiment is performed on the open source infrared image sequence, which can realize the simultaneous positioning and classification of multiple objects, and the processing speed is up to 56 FPS. Experimental results show that this method can effectively identify infrared targets in complex backgrounds in real time.
作者
王倩
张海峰
米娜
尹泽楠
WANG Qian;ZHANG Haifeng;MI Na;YIN Zenan(Big Data Center of State Grid Corporation of China,Beijing 100052,China)
出处
《光学技术》
CSCD
北大核心
2021年第4期483-488,共6页
Optical Technique
基金
国家自然科学基金项目(51577103)。
