Research On Rice Panicle Image Segmentation Based On Multi-scale Features And Attention Mechanism

Along with maize and wheat,rice is the world’s three major staple foods,providing a food source for most of the world’s population.As the reproductive organ of rice,rice panicles are closely related to rice yield.By analyzing the phenotypic characteristics of rice panicles,it can not only provide assistance for the localization and identification of functional gene loci related to rice yield estimation and yield,but also contribute to disease and pest detection,nutritional diagnosis,and growth period determination of rice.Image segmentation technology is the key to automatic acquisition of rice ear phenotypic features.Therefore,the following research work is carried out on the content related to rice ear image segmentation.1.In response to the problem that the existing rice ear image segmentation method based on deep learning cannot effectively cope with the multi-scale problem of rice ear due to using the standard convolution with a single and relatively small kernel scale as the feature extraction operator,a multi-scale convolution-based rice ear image segmentation method LPC-UNet is proposed.Firstly,Lite Pyramid Convolution(LPC),a multi-scale convolution module,is proposed.The module consists of deep separable convolutions to reduce the increase in parameter and computational complexity caused by the addition of new convolutional branches,so as to achieve its lightweight design.Then,the module is used to replace the 3×3 standard convolution in the original U-Net network,so that each layer of the network can have larger and more receptive field to capture more context information and learn multi-scale features from coarse to fine,so as to solve the multi-scale problem of rice ears.The experimental results show that the proposed method can effectively address the multi-scale problem of rice ears and improve the accuracy of rice ear segmentation.2.To address the problem that the LPC module cannot suppress background interference,which causes the network to be prone to mis-segmentation when encountering the mixed color of rice panicles and leaves,a rice panicle image segmentation method combining multi-scale convolution and channel attention mechanism is proposed.Firstly,a Double Branch Squeezeand-Excitation(DBSE)module is designed based on the Squeeze-and-Excitation(SE)module.This module inserts global maximum pooling into the SE module for compensating the information loss caused by global average pooling and maintaining the efficiency of the attention module.To distinguish the importance of the features extracted by the two pooling operations,the two operations were weighted.In addition,the fully connected layer in the SE module is replaced with a one-dimensional convolution to capture the correlation between feature channels,thus reducing its complexity and improving its efficiency.Then,the DBSE module is added to the convolutional branch of LPC to suppress background features and enhance target features during the period of extracting multi-scale features,so as to improve network’s recognition ability for rice ears,and reduce the occurrence of mis-segmentation.The experimental results show that this method can effectively suppress background information interference,reduce the occurrence of mi-segmentation,and further improve the accuracy of rice ear segmentation.3.An automated segmentation system for rice panicle images is developed.The system provides a one-stop service of image preprocessing,segmentation and optimization,which can quickly and accurately segment rice ears and help rice breeders and cultivators to automatically extract and analyze the phenotypic features of rice ears.