一种基于FCOS神经网络的小建筑物目标检测方法

Small Buildings Detection Method Based on FCOS Neural Network

摘要:

提出了一种基于FCOS神经网络的小建筑物目标检测算法,针对FCOS模型在特征提取阶段提取到的小建筑物目标特征较少问题,引入多尺度检测和可变形卷积方式,加强网络对小建筑物目标的特征提取能力,并通过改进后的SGE注意力机制降低特征图中的干扰噪声权重。改进后的网络可以提取到更多的小建筑物目标特征,对环境干扰噪声的鲁棒性更强。在自己搭建的数据集上进行了实验测试,结果表明,在相同环境下网络改进后建筑物的整体检测准确率提升了1.7% ,其中对小建筑物目标提升了3.6%,减少了小建筑物目标漏检、误检的问题。

Abstract:

A small building target detection algorithm is proposed based on FCOS neural network. Aiming at the problem of insufficient target features of small buildings extracted in the feature extraction stage of the FCOS algorithm, multi-scale convolution and deformable convolution were used to improve the ability of the network to extract features of small buildings. And through the improved SGE attention mechanism, the weight of interference noise in the feature map becomes lower. The improved network can extract more target feature information of small buildings and is more robust to noise. Experimental results obtained from the building dataset made by ourselves show that, under the same test environment, the network’s overall detection accuracy (mAP) of buildings under normal, dense and occluded conditions is improved by 1.7%, and that of small buildings is improved by 3.6%, reducing the missing and error detection of small building targets.