Performance and speed optimization of DLV3-CRSNet for semantic segmentation of Chinese cabbage (Brassica pekinensis Rupr.) and weeds

Accurate and efficient recognition of crops and weeds in complex agricultural environments is crucial for promoting intelligent weed management and sustainable farming. Despite DeepLabv3+ showing robust semantic segmentation capabilities, its complex architecture and large number of parameters impede training efficiency and pose deployment difficulties, especially in resource-constrained farming environments. Additionally, it has suboptimal accuracy in segmenting small targets, which limits its ability to identify minor crops and weeds. To address these issues, we propose using ConvNeXt as its backbone, integrating the RepVgg structure, and applying the Sigmoid Linear Unit (SiLU) activation function based on the DeepLabv3+ model (DLV3-CRSNet). Experimental results show that DLV3-CRSNet outperforms DeepLabv3+ by achieving improvements of 23% in Mean Intersection over Union (MIoU), 21% in Mean Pixel Accuracy (MPA), 23% in Average Precision (AP), and 15% in Frames per Second (FPS). Additionally, Floating Point Operations per Second (FLOPS) and inference time (IT) are reduced by 13% and 12%. Compared with Pyramid Scene Parsing Network (PSPNet), Expectation-Maximization Attention Networks (EMANet), Fully Convolutional Networks (FCN), UNet, Efficient Neural Network (ENet), and Segmentation Network (SegNet), DLV3-CRSNet enhances MIoU, MPA, AP, and Accuracy by an average of 27%, 18%, 18%, and 3%, while reducing the number of parameters by 26.84 million (M). Field experiments further confirm that DLV3-CRSNet effectively distinguishes Chinese cabbage (Brassica pekinensis Rupr.) and weeds, achieving recognition rates of 95.31% and 93.6%, respectively.