感受野约束下轻量化神经网络设计

Design of receptive field involved lightweight neural network

针对传统网络模型设计中模型感受野与特征图大小不匹配的问题,该文以待识别的建筑物尺度大小作为先验知识辅助构建模型,在充分融合MobileNetV2和DeepLabV3模型优势的基础上,采用空洞卷积和调整卷积步长搭配的策略,通过调整模型最终感受野与建筑物尺度相契合,构建了感受野约束下的Encoder-Decoder结构网络,即RFNet。利用国产高分二号影像长春市区建筑物数据集对该网络进行了测试,结果表明与参数量占优的模型相比,在交并比接近的情况下,该文的模型参数量降低了74.6%;在与未考虑感受野与特征相契合的网络相比,交并比提高了15.24%,表明本文所设计网络模型的有效性。

Aiming at the problem of mismatch between model receptive field and feature map size in traditional network model design,this paper used the scale of the building to be recognized as a priori knowledge to assist in the construction of the model.On the basis of fully integrating the advantages of MobileNetV2 and DeepLabV3 models,the final perceptual field of the model was adjusted to match the scale of the building by using the strategy of null convolution and adjusting the convolutional step size to construct an Encoder-Decoder structure of RFNet network.The results showed that the number of parameters have been reduced by 74.6%compared with the model with a similar cross-comparison ratio,and the cross-comparison ratio was increased by 15.24%compared with the network without considering the perceptual field and feature fit.