基于特征融合的CenterNet小目标检测方法

CenterNet Small Object Detection Method Based on Feature Fusion

在深度学习目标检测中,小目标指的是待检测图像中覆盖区域较小的一类目标。小目标包含的信息量不足且在一般数据集中数量较少,导致现有的目标检测方法对小目标的检测效果不够理想。针对小目标检测问题,提出一种基于特征融合的CenterNet快速小目标检测方法。该方法根据卷积神经网络不同深度特征的特点,将特征从高到低逐层进行融合,在高分辨率的融合特征上进行预测,提高了模型对小目标的检测能力。同时,针对现有数据集中小目标数量较少问题,提出一种简单有效的数据预处理方法,在训练集中加入高分辨率、低信息量的图像,用其中的大目标帮助模型学习同类、相似小目标特征。实验结果表明,所提出的方法相比于原始CenterNet对小目标的检测能力提升明显。

In deep learning target detection,small object refers to a kind of target with small coverage in the image to be de⁃tected.Small object contains insufficient information and is less in general data sets,which leads to the bad detection effect of exist⁃ing target detection methods.Aiming at the problem of small object detection,a fast small object detection method based on Center⁃Net and feature fusion is proposed.This method uses the characteristics of different depth features of convolutional neural network to fuse the features layer by layer from high to low,and the prediction is carried out on the high-resolution fusion features,which im⁃proves the detection ability of the model for small object.In addition,a simple and effective data preprocessing method is proposed to solve the problem that the quantity of small targets in the existing data set is low.High resolution and low information images are added to the training set,and the large object are used to improve the detection rate of small object.Experimental results show that the detection ability of the proposed method is significantly improved compared with the original CenterNet.