基于白平衡特征增强的秸秆目标分割方法  

Straw Target Segmentation Method Based on White Balance Feature Enhancement

作　　者：姜含露 王飞云 潘宇轩 刘阳春[1,2] 汪凤珠 周利明[1,2] 吕程序 JIANG Hanlu;WANG Feiyun;PAN Yuxuan;LIU Yangchun;WANG Fengzhu;ZHOU Liming;LÜChengxu(Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd.,Beijing 100083,China;State Key Laboratory of Agricultural Equipment Technology,Beijing 100083,China)

机构地区：[1]中国农业机械化科学研究院集团有限公司,北京100083 [2]农业装备技术全国重点实验室,北京100083

出　　处：《农业机械学报》2024年第S1期92-100,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：北京市科技新星计划项目(20220484066)

摘　　要：针对图像中多种复杂环境干扰,本文结合秸秆在东北黑土地上具有的颜色优势,提出一种基于白平衡特征增强的秸秆目标分割模型。DLv3+/CPM/SEM模型采用编解码结构,在DLv3+模型基础上融合了颜色感知模块CPM与空间增强模块SEM,利用白平衡技术提高秸秆目标在图中的对比度使其在多种干扰因素的影响下仍能保持对秸秆目标检测的准确率。其中编码部分利用残差网络组成双分支特征提取结构,通过全反射算法增强秸秆颜色特征的同时,消除光线条件对图像颜色显现的干扰,双分支特征通过级联的感知方式并入颜色感知模块CPM,以加强补色的形式对图像中偏色严重的秸秆进行多层级的颜色特征增强,从而提取准确的秸秆特征表达;解码部分将整合的特征代入具有ASPP的解码模型中,加入空间增强模块SEM提高秸秆和农田背景的区分度,优化秸秆目标分割模型性能。经试验验证,提出的DLv3+/CPM/SEM改进模型在准确率和MIoU的模型整体评价指标上都高于其它对比模型,在不同光源条件、秸秆长度、垄沟深浅和土块大小的干扰条件下均有较好的分割效果,同时结合距离划分结果后,对非单一农田背景的秸秆监测图像的覆盖率计算精度更为准确。In response to the interference of various complex environments in images,this paper proposes a straw target segmentation model based on white balance feature enhancement,taking into account the color advantage of straw on black soil in Northeast China.The DLv3+/CPM/SEM model adopts an encoder decoder structure,which integrates the color perception module CPM and spatial enhancement module SEM on the basis of the DLv3+model.The white balance technology is used to improve the contrast of straw targets in the image,so that they can still maintain the accuracy of straw target detection under the influence of various interference factors.The encoding part utilized a residual network to form a dual branch feature extraction structure,which enhances the color features of straw through total reflection algorithm while eliminating the interference of light conditions on the color display of the image.The dual branch features are merged into the color perception module CPM through a cascaded perception method to enhance the color features of straw with severe color cast in the image at multiple levels in the form of reinforced complementary colors,thereby extracting accurate straw feature expressions.The decoding part incorporates the integrated features into the decoding model with ASPP,and adds a spatial enhancement module SEM to improve the discrimination between straw and farmland background,optimizing the performance of the straw target segmentation model.Through experimental verification,the improved DLv3+/CPM/SEM model proposed in this paper has higher accuracy and overall evaluation indicators of MloU than other comparative model models.lt has good segmentation effects under different light source conditions,straw length,ridge depth,and soil block size interference conditions.At the same time,combined with the distance segmentation results,the coverage calculation accuracy of straw monitoring images with nonsingle farmland backgrounds is more accurate.

关 键 词：秸秆 分割 DLv3+ 空间增强模块 颜色感知模块 

分 类 号：S141[农业科学—肥料学] TP391.41[农业科学—农业基础科学]