基于滤波网络的多余物状态识别研究

Research on Remainer State Identification Based on Filtering Network Method

摘要:

在航天产品的研发制造过程中,多余物控制至关重要,尤其是多余物状态识别的环节,其核心在于如何有效提取高噪声图像中的局部特征。现有方法尚未针对多余物场景进行有效建模,通用的视觉模型往往会对噪声进行过度拟合,而无法有效滤除噪声信号。为此,提出了一种可学习滤波感知机。该感知机通过采用一种可学习滤波器来替代繁琐的自注意力机制,用以学习空间位置的交互作用信息。随后,引入了频谱掩模用于频域分量特征的抽取,以学习不同频段内的重点信息。实验结果表明,该方法在多余物识别中取得了96.7%的准确率,优于基于卷积和自注意力的模型,并且具有更低的计算复杂度。

Abstract:

The remainder control is crucial to the development and manufacturing of aerospace products, and the remainders state recognition is an important part of it. The key which is to effectively extract local features in high noise pictures. However, existing methods have not been modeled well specifically for remainder scenes, and generic vision models are prone to overfitting the noise, making it difficult to filter the noisy signals effectively. To solve this problem, this paper proposes a learnable Filter Network, which replaces the heavy self-attention mechanism by a learnable filter which is used to learn spatial location interaction information. And then incorporates a mask for frequency domain component feature extraction to learn the emphasis information of different frequency bands. It is experimentally demonstrated that this method works better in remainder recognition scenarios, outperforms the convolution and self-attention models, and has better time complexity.