Adaptive multiple subtraction is an important step for the success of multiple removal after multiple prediction. Generally, the traditional method uses a 2D matching filter to combine the predicted multiples to match with the original data directly. Due to the complicated mismatches between the predicted multiples and true multiples, multiples may be removed aggressively with damaging primaries and vice versa for the traditional method. Especially in complex media, how to balance multiple removal and primary preservation is very important. In this paper we propose to use multi feature-gathers of the predicted multiples for adaptive multiple subtraction. The feature of the predicted multiples is extracted by the convolutional neural network with the predicted multiples as the input and the original data as the output. The multi feature-gathers of the predicted multiples contain more prediction information than the predicted multiples themselves. Therefore, the multi feature-gathers combined by a 3D matching filter can better match with the true multiples than the predicted multiples themselves combined by a 2D matching filter. Synthetic and field data examples demonstrate that the proposed method can better balance multiple removal and primary preservation than the traditional method.

中文翻译：

---

基于卷积神经网络的自适应多次相减特征提取

自适应多次减法是多次预测后成功去除多次的重要步骤。通常，传统方法使用2D匹配滤波器来组合预测倍数以直接与原始数据匹配。由于预测倍数和真实倍数之间存在复杂的不匹配，对于传统方法，倍数可能会因具有破坏性的原色而被主动删除，反之亦然。特别是在复杂的介质中，如何平衡多次清除和一次保存非常重要。在本文中，我们建议使用预测倍数的多重特征集进行自适应多重减法。通过卷积神经网络提取预测倍数的特征，将预测倍数作为输入，将原始数据作为输出。预测倍数的多重特征集合包含比预测倍数本身更多的预测信息。因此，与由2D匹配滤波器组合的预测倍数本身相比，由3D匹配滤波器组合的多特征集可以更好地与真实倍数匹配。综合和现场数据实例表明，与传统方法相比，该方法可以更好地平衡多次清除和一次保存。