In recent years, a large number of quantization schemes have been proposed for compressing convolutional neural networks (CNN). However, most of them have the following problems: 1) when there are outliers in the weight, post-training quantization cannot obtain the ideal effect, and the accuracy loss is unavoidable; 2) quantizing the non-maximum suppression (NMS) stage of oriented object detection networks is non-trivial so that such networks are difficult to deploy on edge computing devices that only support integer operations. In this letter, we propose the outlier-aware quantization (OAQ) to boost the robustness of the post-training quantization method. Besides, we design a multilayer perceptron network to approximate the intersection-over-union (IoU) of rotated boxes, making the NMS stage can be deployed on integer-arithmetic-only devices. The experiment results demonstrate that our solution outperforms the widely used post-training quantization method. Meanwhile, to the best of our knowledge, this is the first study that focuses on the optimization and quantization of the NMS stage of oriented object detection networks.

中文翻译：

---

通过异常值感知量化和 IoU 近似量化面向对象检测网络

近年来，已经提出了大量用于压缩卷积神经网络（CNN）的量化方案。然而，它们大多存在以下问题：1）当权重存在异常值时，训练后量化无法获得理想效果，精度损失不可避免；2) 量化面向对象检测网络的非极大值抑制 (NMS) 阶段非常重要，因此此类网络难以部署在仅支持整数运算的边缘计算设备上。在这封信中，我们提出了异常值感知量化 (OAQ) 以提高训练后量化方法的鲁棒性。此外，我们设计了一个多层感知器网络来近似旋转框的交集（IoU），使 NMS 阶段可以部署在仅整数算术的设备上。实验结果表明，我们的解决方案优于广泛使用的训练后量化方法。同时，据我们所知，这是第一项侧重于面向对象检测网络的 NMS 阶段的优化和量化的研究。