基于多模态深度学习的图像序列弱小目标识别

Image Sequence Weak and Small Target Recognition Based on Multi-Modal Deep Learning

针对现有图像序列弱小目标识别存在图像特征学习不全面、训练样本较大,导致对相似物体识别率及准确率较低的问题,提出基于多模态深度学习的图像序列弱小目标识别。利用弱小目标与背景之间的相关性对单帧图像进行背景抑制,得到目标和高频噪声,对图像做目标分割处理,剔除高频噪声。在此基础上,使用加入稀疏性约束的自编码器不断调节其自身参数,压缩输入信息,提取有用的输入特征,训练出最佳唯一向量,使用优化的CNN深度学习模型完成弱小目标识别。实验结果表明,所提方法能够在不依赖大量识别训练的情况下,始终保持较高的识别率,最大识别率为99.21%,优于传统方法。

Traditional image sequence weak small target recognition has defects, such as incomplete image feature learning and large training samples, thus leading to the problem of low recognition rate and accuracy for similar objects. Therefore, an image sequence weak and small target recognition based on multi-modal deep learning was proposed in this work. The background of a single frame image was constrained by using the correlation between the weak and small target and the background. Then, the target and high-frequency noise were obtained, and the image was segmented to remove the high-frequency noise. Then, the self-encoder with sparsity constraint was used to adjust its parameters. The useful input features were extracted by compressing the input information. The useful input features were extracted to train the best unique vector, thus achieving the small target recognition. The results show that this method is superior to the traditional methods, and has a high recognition rate. The maximum recognition rate is up to 99.21%.