基于改进U-Net模型的近海水产养殖池塘信息提取

Information extraction from offshore aquaculture ponds based on improved U-Net model

针对传统卷积神经网络算法在复杂环境背景下泛化能力弱、提取精度低的问题,提出了一种快速、准确的近海养殖池塘自动提取模型。该模型在U-Net模型的基础上进行改进,用新提出的多尺度特征提取结构(DC)取代了U-Net模型的传统卷积层。其中,DC结构融合了Inception模块和空洞残差模块,增强了模型提取养殖池塘特征的能力,降低模型训练过拟合风险,有效减少图像背景信息干扰。实验结果表明,改进U-Net模型的精确率、召回率、交并比、F1分数分别为91.73%、90.47%、91.12%、89.91%,优于其他的对比模型。该研究能够有效提高养殖池塘提取精度,实现养殖池塘快速、准确的监测。

A fast and accurate model for automatic extraction of offshore aquaculture ponds is proposed to address the problems of weak generalisation and low extraction accuracy of traditional convolutional neural network algorithms in complex environmental contexts.The model improves on the U-Net model by replacing the traditional convolutional layers of the U-Net model with a newly proposed multi-scale feature extraction structure(DC).The DC structure incorporates the Inception module and the null residual module to enhance the ability of the model to extract farm pond features,reduce the risk of overfitting the model training and effectively reduce the interference of image background information.The experimental results show that the accuracy,recall,cross-merge ratio and F1 score of the improved U-Net model are 91.73%,90.47%,91.12%and 89.91%respectively,which are bet-ter than the other comparison models.This study can effectively improve the extraction accuracy of farming ponds and achieve fast and accurate monitoring of farming ponds.