基于卷积评价及对抗网络的花粉、孢子图像增广算法

Augmented algorithm for pollen and spore images based on convolution evaluation and pix2pix network

针对花粉、孢子图像特征复杂,样本稀缺及种类繁多制约图像检测识别效果的问题,建立基于自适应阈值分割的pix2pix图像增广模型。首先基于卷积评价改进自适应阈值分割算法,择优选取语义分割图像;其次构建pix2pix图像增广模型,将语义分割图像和原始图像建立标签映射用于模型训练,根据语义分割图像生成仿真图像,扩充样本数据集。结果表明,以149种花粉、孢子图像为样本,通过图像增广模型生成的花粉、孢子图像整体相似度达到85.40%;图像增广前Faster RCNN、YOLOv3检测模型的检测精准率分别为86.18%、85.64%,使用增广后的样本训练模型,检测精准率分别达到92.16%、90.57%,提升5.98个百分点和4.93个百分点。

Aiming at the problems such as complex image features of pollens and spores,scarcity of image samples,restricted detection and recognition effects of various images of pollens and spores,a pix2pix image augmentation model based on adaptive threshold segmentation was built.Firstly,the adaptive threshold segmentation algorithm was improved based on convolution evaluation to select the optimal semantic segmentation images of pollens and spores.Secondly,the pix2pix image augmentation model was constructed,the semantic segmentation images and the original images were used to establish label mapping for model training,and emulational pollen and spore images were generated based on semantic segmentation images to extend sample dataset.The results showed that,the overall similarity of 149 pollen and spore images generated by the image augmentation model reached 85.40%.Before image enlargement,the detection accuracies of Faster RCNN and YOLOv3 detection models were 86.18%and 85.64%,respectively.After using the enlarged sample training model,the accuracies reached 92.16%and 90.57%,which increased by 5.98 and 4.93 percentages.