Maize crop row recognition algorithm based on improved UNet network

Abstract

Aiming at the problem that it is difficult to identify maize crop row centerlines in complex farmland environments such as high weeds, row breaking, and leaf adhesion under different growth periods, this study proposes a centerline detection algorithm based on a improved UNet network. The UNet network – a traditional semantic segmentation network – was enhanced to create the Atrous Spatial Pyramid Pooling UNet (ASPP-UNet) network for maize crop row and background segmentation, and the improved vertical projection method was subsequently employed to measure the crop rows’ feature points. Finally, the least squares method was used to fit the centerlines. Experimental results yielded the Mean Intersection Over Union, Mean Pixel Accuracy, Mean Precision, and Mean Recall metrics of ASPP-UNet network to be 83.23%, 90.18%, 91.79%, and 90.18% respectively. These figures represent respective increases of 10.03%, 11.86%, 9.43% and 11.24% compared to the Fully Convolutional Network (FCN), and 7.80%, 5.52%, 2.71%, and 5.52% compared to the UNet. Furthermore, the average fitting time and angle error of the improved vertical projection method combined with the least square method were reduced to 66 ms and 4.37°, compared to 80 ms and 6.12°in the traditional vertical projection method, and 86 ms and 5.67°in the left and right edge centerline method. Likewise, the accuracy of proposed method increased to 92.59%, compared to 87.21% and 90.16% of the two aforementioned methods, respectively. Therefore, the proposed method successfully meets the accuracy and real-time demands of agricultural robot vision navigation, and can function effectively under varying environmental pressures.