Optimization of Intelligent Plant Cultivation Robot System in Object Detection

Abstract

Intelligent plant cultivation robot is an emerging device in the field of intelligent plant cultivation. The device can solve many problems that cannot be solved by Non-intelligent plant cultivation device in the past. Moreover, compared with other devices, intelligent plant cultivation robot is more convenient, low-cost and efficient. And in the process of intelligent robot plant breeding work, visual recognition is one of the most important step. In this paper, on the steps of using visual identification module in the traditional yolov3 visual identification algorithm is improved, in does not affect yolov3 visual identification algorithm under the condition of its running speed. In fact, due to its relatively low recognition accuracy, the high rate of error rate and empty selected situation, we have carried on corresponding optimizations and innovations such as,Attention mechanisim namely SE(squeeze and excitation) module, which can help the network to focus on more inportant fearture and enhancing the whole performance(higher accuracy, lower error rate) of the network. In addition, we changed the original backbone Darknet53 to GhostNet, which greatly reduced the parameters(also FLOPs) during the network inference and improved the precision to a certain extent. The loss function is optimized,namely the traditional IOU is changed into CIOU, which reduces the empty selection rate and improves recognition accuracy (2%). Experiments show that our method is better(better precision, recall, F1 and less inference time) than the original method both on the VOC dataset and our self-made plant dataset.