一种航空影像建筑物检测的轻量化CNN建模方法  

A Lightweight CNN Modeling Method for Building Detection from Aerial Images

作　　者：甘文祥 张远谊 李欣园 GAN Wenxiang;ZHANG Yuanyi;LI Xinyuan(School of Geography and Information Engineering,China University of Geosciences(Wuhan),Wuhan 430074,China)

机构地区：[1]中国地质大学(武汉)地理与信息工程学院,湖北武汉430074

出　　处：《地理空间信息》2023年第6期24-27,共4页Geospatial Information

基　　金：国家自然科学基金资助项目(41601506)。

摘　　要：以卷积神经网络为代表的深度学习方法大幅提高了遥感影像建筑物自动检测精度,但由于建筑物复杂多样,为了提取区分能力更强的图像特征,现有卷积神经网络方法往往倾向于构建层次复杂、参数庞大的深度模型。这使得模型的存储和内存开销都较高、检测速率也容易受到影响,一定程度上造成在移动设备平台或灾害应急等场合的应用受限。针对此问题,提出一种用于航空影像建筑物检测的轻量化卷积神经网络建模方法,采用深度可分离卷积方法对复杂网络进行简化,大幅减少了计算量,并较好地维持了原有精度。实验表明新方法相比改进前,在计算量和参数量分别减少86%和87%、训练时间缩短10%的情况下,建筑物检测的精度仅降低3%。In recent years,represented by convolutional neural network(CNN)method,the deep learning method has greatly improved the auto-matic detection accuracy of buildings in remote sensing images.However,due to the complexity and diversity of buildings,in order to construct and extract image features with stronger distinguishing ability,existing CNN methods tend to construct deep models with complex layers and large parameters.This makes high storage and memory overhead of the model,and the detection rate easy to be affected,which to some extent limits the application in mobile device platform or disaster emergency.To solve this problem,we proposed a lightweight CNN modeling method for building detection from aerial images.We used the depthwise separable convolutional method to simplify the complex network,which could greatly reduce the computation and maintain the original accuracy.The test results show that compared with the method before the improvement,the accuracy of building detection is only reduced by 3%when the calculation and reference quantity are reduced by 86%and 87%respectively and the training time is shortened by 10%.

关 键 词：航空影像 建筑物检测 CNN 深度可分离卷积 轻量化网络 

分 类 号：P237[天文地球—摄影测量与遥感]