Sugarcane mechanized planting technology consists of seed preparation and field planting. This study aims at the issues of easy damage to the seeds during the operation of the automatic cutting machine for single-bud segment sugarcane, lack of intelligent seed selection and calibration technology, low recognition accuracy, and the need for manual feeding of the planting machine’s seed meter which leads to seed leakage. This study, based on machine vision and deep learning, optimizes the seed calibration method and proposes an improved YoloV5-STD target detection algorithm to improve the recognition accuracy of seed characteristics and optimize the overall engineering structure. For the planting machine, a new type of hopper for the seed meter is designed using natural rubber as the base material mixed with polystyrene, and the flexible automatic seed metering mechanism is analyzed to achieve automatic feeding and seed metering. Test assessment indicators were formulated based on the enterprise standards of the Institute of Agricultural Machinery Research, Chinese Academy of Tropical Agricultural Sciences. Experimental results show that the recognition accuracy of the 2DZ-2 type single-bud segment intelligent cutting machine is ≥95%, the bud injury rate is <1.8%, the qualified rate of cutting is 95.8%, and the single-channel cutting efficiency is 64 buds/min. The 2CZD-2C type single-bud segment planter has a planting qualification rate of 96.6%, a planting efficiency of 208 buds/min, and a seed leakage rate of <2.1%.

In order to avoid the uneven phenomenon of sugarcane planting, such as seed missing and reseeding, the computer vision technology was applied to the intelligent identification of sugarcane varieties with single-bud segment, and the design idea of rapid detection of sugarcane planting distribution was proposed in this study. With sugarcane species with single-bud segment as the research object, the sugarcane species distribution image was acquired, and LabelImg was used for image annotation and format conversion to build the YOLOv5s target detection model. On the basis of depth-separable convolution, SE module is spliced to obtain the weights of extracted features and extract key features of input feature map. By adding regularization to constrain the BN layer coefficient, sparse regularization is carried out on the BN layer to reduce the network input size and improve the model training speed. On this basis, 600 rounds of iterative training were carried out to complete the target recognition of sugarcane species characteristics in single-bud segment. The results showed that the recognition accuracy, mAP value, and Recall value of YOLOv5s single-bud segment target detection model are 98.95%, 98.89%, and 98.69%, and the loss value converges in advance between 0-0.02. The results showed that YOLOv5s could effectively detect and identify sugarcane seeds with single-bud segment during field planting, which lays a foundation for promoting precise and intelligent sugarcane planting.