关键词: 植物叶片, U2—DSCNet, 语义分割, 残差连接, 深度可分离卷积

Abstract: In order to enhance the accuracy and effectiveness of plant leaf segmentation algorithms in complex backgrounds and reduce the computational load of convolution, an improved U2—Net semantic segmentation model, U2—DSCNet, is proposed. This model introduces Depthwise separable convolution based on the RSU residual structure of U2—Net. The RSU unit of U2—Net is replaced by using the DSC—RSU residual module toachieve an improved U2—Net semantic segmentation model, U2—DSCNet. The model consists of encoder, decoder, and feature fusion. The encoder has 6 encoding modules (En_1—En_6), and the decoder has 5 decoding modules (De_1—De_5). Subsequently, feature fusion is performed on the output images from the 5 decoders, ultimately obtaining a feature map that integrates semantic information from different scales for model training. Model validation experiments were conducted on a test set and images collected in natural environments, and the results were compared with those of algorithms such as FCN, SegNet, U—Net, and U2—Net. By using precision rate, recall rate, Fβ, and IoU as evaluation metrics, U2—DSCNet achieved results of 0.952, 0.956, 0.952, and 80.3 in these four metrics, showing significant improvements compared to the other segmentation algorithms. Moreover, the model size and training efficiency are also better than those of U2—Net, with a model size of 137 MB and a training time of 0.72 s. The experimental results demonstrate that this model not only achieves high segmentation accuracy but also possesses good generalization capabilities.

Key words: plant leaves, U2—DSCNet, semantic segmentation, residual connection, deep separable convolution