Several approaches to the use of neural networks for object detection on spatially inhomogeneous backgrounds are considered. A method for constructing a classifier for object detection directly from observed fragments has been developed. An approach consisting of a combination of matched linear filtering and convolutional neural networks is proposed. It is shown that this approach reduces the false alarm probability while maintaining the object detection probability.

中文翻译：

---

使用神经网络在空间非均匀背景上进行目标检测

考虑了使用神经网络在空间不均匀背景上进行物体检测的几种方法。已经开发了一种直接从观察到的片段构建用于对象检测的分类器的方法。提出了一种由匹配线性过滤和卷积神经网络组合组成的方法。结果表明，该方法在保持目标检测概率的同时降低了误报概率。