Lightweight crop and weed recognition method based on imporved DeepLabv3+

Abstract

To achieve field crop and weed recognition on devices with limited storage resources and computational capabilities, a lightweight semantic segmentation network based on improved DeepLabv3+ is proposed. Firstly, MobileNet v2 is used as the feature extraction backbone for DeepLabv3+, where the residual modules are replaced with dual-branch residual modules and the last two convolutional layers are removed to reduce the model parameters. Secondly, group-wise pointwise convolution is introduced in the Atrous Spatial Pyramid Pooling module, replacing standard convolutions with depthwise dilated convolutions, and performing multi-scale feature fusion on the convolved feature maps to enhance the extraction of deep features for crops and weeds. Lastly, the original non-linear activation functions are replaced with the Leaky ReLU activation function to improve segmentation accuracy. Experimental results show that the improved DeepLabv3+ achieves an mIOU (Mean Intersection over Union) of 86.75% with only 0.69M parameters, and achieves an FPS (Frames Per Second) of 98. Compared to the original DeepLabv3+ and three typical lightweight semantic segmentation networks, it has the lowest parameter count and the highest segmentation accuracy among the compared lightweight networks.

Keywords

crop and weed identification / lightweight / semantic segmentation / Deeplabv3+ / MobileNet v2 / multi-scale feature fusion