空间目标序列图像识别技术

针对远距/近距空间目标成像的特点,提出一种基于序列图像的多尺度自动目标识别(ATR)方案.该方案综合利用目标的尺度变化、姿态变化及图像特征信息,分别构建多尺度目标分类器、姿态判别器,并估计目标识别结果可信度、相邻帧姿态变化的权重以及目标尺度权重;根据当前帧和上一帧的识别结果,进行目标类别的融合判别.对STK产生的10类仿真空间目标进行测试,试验结果表明:对远距空间目标,由于目标像素少,仅用单帧图像的识别率低,合理利用目标序列图像包含的信息,可有效提高目标识别率.

Automatic space target recognition based on sequential images

According to the space target imaging characteristics in long/short range,an automatic target recognition method with multiple scales is proposed based on sequential images. By integrating varying scale,varying attitude and image features,multi-scale classifiers were constructed with support vector machines,and attitude identification was conducted through RBF neural networks. Then the estimation on the confidence of recognition results,the weight of varying attitude between two adjacent images,and the weight of target scale was discussed respectively. And the final recognition result was obtained by fusion judgment,according to the current recognition result and the previous recognition result. Tests on 10 classes of video data simulated by STK show that the proposed method is effective. For space targets in long range,owing to small pixels,the recognition accuracy is low with only current image,while the high recognition accuracy can be achieved using sequential images.