一种改进残差网络的农田识别算法

A Farmland Recognition Algorithm Based on Improved Residual Network

农田识别是保证农机自动化作业不越界的前提与基础,为了给农机自动化作业提供环境信息,提出一种基于改进残差网络的农田识别算法。该算法以残差网络(ResNet18)为基础主干网络进行改进,在残差结构的基础上加入通道注意力机制,增强了有效特征的权重,降低了噪声等无用特征的影响,提高了模型的特征提取能力以及鲁棒性;结合标签平滑的训练方式以及学习率余弦退火衰减算法,使用Silu激活函数代替原残差网络中的Relu激活函数,优化了模型的收敛效果。实验结果表明,改进后的残差网络模型对农田、草地、建筑、裸地以及荒漠的识别准确率可达到98.92%,相比于原ResNet18模型的准确率提高了3.8%,证明了该模型的优越性。

Farmland identification is the premise and foundation to ensure that the automatic operation of agricultural machinery does not cross the boundary.In order to provide environmental information for the automatic operation of agricultural machinery,a farmland identification algorithm based on improved residual network is proposed.The algorithm is improved based on the residual network(ResNet18),and the channel attention mechanism is added on the basis of the residual structure,which enhances the weight of effective features,reduces the influence of useless features such as noise,and improves the characteristics extraction ability and robustness of the model.Combined with the label smoothing training method and the learning rate cosine annealing decay algorithm,the Silu activation function is used to replace the Relu acti‐vation function in the original residual network,which optimizes the convergence effect of the model.The experimental results show that the improved residual network model has a recognition accuracy rate of 98.92%for farmland,grassland,buildings,bare land and desert,which is 3.8%higher than the original ResNet18 model,which proves the superiority of this model.