Straw Target Segmentation Method Based on White Balance Feature Enhancement
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 MIoU than other comparative model models.It 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.
国家科技期刊平台
NSTL
万方数据
