Recognition method of maize crop rows at the seedling stage based on MS-ERFNet model

Abstract

It is crucial to carry out weeding-cultivating and pesticide application for maize during the seedling stage since weeds, diseases, and pests in the field might negatively impact the growth of maize at the seedling stage. Aiming at the problems of traditional machine vision methods to identify seedling maize crop rows which are easily affected by light and weeds, low recognition accuracy and poor real-time performance, an improved Multi-Scale Efficient Residual Factorized ConvNet (MS-ERFNet) model was proposed for the recognition of seedling maize crop rows. The camera was used to collect maize crop rows images at the seedling stage, which were expanded using image enhancement to create the corresponding data set. A Multi-Scale feature extraction module (MS module) was proposed, with 3 × 3 convolution as the main branch, and parallel connection with 1 × 1 convolution and 5 × 5 depth separable convolution, and then connecting it with two 3 × 3 asymmetric group convolutions in series to realize the extraction of multi-scale features of crop rows. Meanwhile, Recurrent Feature-Shift Aggregator (RESA) module was added to the ERFNet model to enhance the model's ability to extract objects with extensibility. Finally, the crop rows centerlines were fitted utilizing the Least Squares Method (LSM) after the crop row center points had been located and classified. On the test set of seedling maize crop rows, the recognition comparison experiments were carried out on the model of DeepLabv3+, Efficient Nerual Network (ENet), ERFNet, Fully Convolutional Network (FCN-8 s), U-Net, and MS-ERFNet. The experimental results showed that the mean intersection over union (mIoU) and the pixel accuracy (PA) of the MS-ERFNet model were 93.40% and 97.54%, respectively, which were higher than other models. The MS-ERFNet model has 4.68 M network parameters, and the segmentation speed is 27.94 frames/s, and the speed of processing for images recognition of seedling maize crop rows using the method in this paper is 15.26 frames/s, which satisfies the real-time requirements of maize crop rows recognition. The MS-ERFNet model proposed in this paper provides technical support for the realization of automatic field navigation of agricultural machines.