图像阴影去除对Mask R-CNN-识别效果提升的研究

Research on Mask R-CNN recognition rate improvement by image shadow removal

首先,针对西瓜叶片阴影,提出了基于YCbCr颜色空间的阴影去除方法;其次,建立三个不同的训练集:未去除阴影的西瓜叶片图像训练集D1、阴影去除的西瓜叶片图像训练集D2,一半未去除阴影一半阴影去除的西瓜叶片图像训练集D3;最后,利用目标分割算法Mask R-CNN训练出三个识别模型M1、M2和M3。实验表明:模型M1的识别效果最低,模型M3适中,模型M2的识别效果最佳。

A method of watermelon leaf shadow removal is proposed based on the YCbCr color space.Three different training sets were established.The training set D1 of the watermelon leaf image without shadow removal.The training set D2 of the watermelon leaf image with shadow removal.The training set D3 of the watermelon leaf image with half-shadow removal.The target segmentation algorithm Mask R-CNN is used to train three recognition models M1,M2 and M3.The experimental results show that M1 has the lowest recognition effect,model M3 has a moderate effect,and model M2 has the best recognition effect.