基于U-Net和特征金字塔网络的秸秆覆盖率计算方法  

Calculation Method of Straw Coverage Based on U-Net Network and Feature Pyramid Network

作　　者：马钦[1,2] 万传峰 卫建 汪玮韬 吴才聪 MA Qin;WAN Chuanfeng;WEI Jian;WANG Weitao;WU Caicong(College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China;Key Laboratory of Agricultural Machinery Monitoring and Big Data Application,Ministry of Agriculture and Rural Affairs,Beijing 100083,China)

机构地区：[1]中国农业大学信息与电气工程学院,北京100083 [2]农业农村部农机作业监测与大数据应用重点实验室,北京100083

出　　处：《农业机械学报》2023年第1期224-234,共11页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家发展改革委员会综合数据服务系统基础平台建设项目(JZNYYY001)。

摘　　要：针对田间秸秆覆盖分散、秸秆形态多样,细碎秸秆识别困难,传统图像识别方法易受光照、阴影等因素干扰等问题,本文以黑龙江省齐齐哈尔市龙江县为研究地点,构建田间秸秆图像数据集;对图像进行裁剪、标注后,构建了以U-Net为基础网络的秸秆检测模型。将编码阶段的网络结构换成ResNet34的前4层作为特征提取器,增加模型的复杂度,增强秸秆特征的提取;为增强秸秆边缘识别,在最高语义信息层对深层特征图使用多分支非对称空洞卷积块(Multibranch asymmetric dilated convolutional block, MADC Block)提取多尺度的图像特征;为增加细碎秸秆的检测能力,在跳跃连接阶段使用密集特征图金字塔网络(Dense feature pyramid networks, DFPN)进行低层特征图和高层特征图的信息融合,利用特征图对应秸秆图像中感受野的不同,解决秸秆形态多样的问题;为避免秸秆特征图在上采样时的无效计算,解码阶段使用快速上卷积块(Fast up-convolution block, FUC Block)进行上采样,避免秸秆特征图在上采样时的无效计算。试验表明,本文算法在车载相机采集到的秸秆图像数据集上平均交并比为84.78%,相比U-Net提高2.59个百分点,该网络对于640像素×480像素的图像平均处理时间低于3 ms,符合作业检测时的时间复杂度要求,算法在一定程度上改善了阴影区域秸秆的识别问题,提高了细碎秸秆的识别能力。In view of the scattered straw mulching in the field, the various straw shapes, the difficulty in identifying the fine straw, and the traditional image recognition methods are disturbed by factors such as light and shadow easily. Taking Longjiang County, Qiqihar City, Heilongjiang Province as the research site, a field straw image dataset was constructed. After cropping and labeling the image, a straw detection model based on U-Net network was constructed. Changing the network structure of the coding stage to the first four layers of ResNet34 as the feature extractor, the complexity of the model was increased and the extraction of straw features was enhanced. In order to enhance the detailed identification of straw edges, the multibranch asymmetric dilated convolutional block(MADC Block) was used to extract multi-scale image features on the deep feature map at the highest semantic information layer. In order to increase the detection ability of fine straws, dense feature pyramid networks(DFPN) were used in the skip connection stage to perform information fusion of low-level feature maps and high-level feature maps. Using the feature map to correspond to the difference of the receptive fields in the straw image, the problem of variety of straw shapes was solved. In order to avoid the invalid calculation of straw feature map during upsampling, the decoding stage used fast up-convolution block(FUC Block) was used for upsampling. Experiments result showed that the average intersection ratio of the algorithm on the straw image dataset collected by the vehicle camera was 84.78%, which was 2.59 percentage points higher than that of U-Net. The average processing time of the network for images with a size of 640 pixels×480 pixels was less than 3 ms. Compared with manual measurement, the error was less than 5%, which met the time complexity requirements of operation detection. The algorithm can improve the identification of straw in the shadow area to a certain extent, and improve the identification ability of fine straw.

关 键 词：秸秆检测 计算机视觉 非对称空洞卷积 特征图金字塔网络 

分 类 号：S23[农业科学—农业机械化工程] TP391[农业科学—农业工程]