Research On Field Weed Segmentation Method Based On Channel Attention Mechanism

Field weeds are one of the important factors affecting crop yield reduction.At present,pesticide spraying has become the main means of weed control in agriculture.However,farmers generally use untargeted pesticide spraying method,completely ignoring the distribution of weeds in actual agricultural production.This not only leads to low pesticide utilization rate,environmental pollution and crop chemical residues,but also causes safety problems of agricultural products.With the development of electronic information technology,machine vision combined with image processing technology has become an effective method for accurate and real-time weed and crop detection in the field,providing valuable information for weed management in specific locations.However,the traditional digital image processing technology applied to the identification of crops and weeds is easily affected by illumination,complex background,artificial feature selection and other factors,which leads to the reduction of the recognition accuracy.In recent years,with the rapid development of computer vision,the use of convolutional neural networks to achieve accurate segmentation of crops and weeds has become the future development trend of precision agriculture.To address the shortcomings of existing methods,this paper combines the knowledge of convolutional neural network with the feature extraction ability of image semantic segmentation model to design a more efficient network model for real-time weed segmentation in the field.The main work of our paper is organized as follows:(1)Due to the small agricultural dataset,training a deeper CNN network model can easily lead to overfitting,while the shallow CNN structure cannot extract the deep abstract features well.To address the above problems,this paper designs a network model with an efficient coding and decoding structure.The pixel location information is recorded by indexing to solve the problem of detail loss due to the maximum pooling operation;in addition,the features extracted in the encoding stage are fused with the output features upsampled in the decoding stage through a jump link structure to refine the segmentation boundary.(2)Field operations have certain requirements for the real-time performance of agricultural robots.Complex network models require more memory space and computational resources due to the huge number of parameters,while usually embedded devices have limited computational resources for larger scale computations.Therefore,reducing the number of parameters and computation of the model is one of the important factors to solve the implementation operation of agricultural robots.The model structure proposed in this paper uses separable convolution,which greatly reduces the number of model parameters and saves computational resources without basically losing the accuracy of the convolution operation.(3)To avoid the problem of feature information loss due to dimensionality reduction,cross-channel interaction allows the network to focus on the key information of each channel feature and enhances the expressiveness of the network.In this paper,an efficient channel attention mechanism is used among the modules to improve the segmentation performance.Finally,relevant experiments were conducted on corn seedling small sample dataset and carrot small sample dataset,and the model has high segmentation accuracy compared with mainstream methods,and its performance is further analyzed to verify that the model has good real-time performance.