Research On Real-time Detection Module Of Citrus Based On Deep Learning

Citrus is one of the most common fruits in our country.Especially in recent years,with the gradual improvement of material living standards,social demand gradually tends to be diversified,and people’s demand for citrus increases sharply.Citrus is also one of the main agricultural products all over the world,as an agricultural country,China has widely planted citrus in all parts of the country with huge output.Unfortunately,the development of China’s citrus industry is severely restricted by the backward orchard management technology and the low degree of automation,information technology and intelligence.With the rapid development of artificial intelligence,the technology of computer vision based on the deep learning is widely used in agricultural production,and plays an important role in it.It has an amazing performance in all aspects,and can provide a lot of efficient production guidance for agricultural production,but the cost is very high,it goes against the aim of precision agriculture in China advocate,namely "low cost,high return”.In this paper,the object detection algorithm called YOLOV3 was improved for citrus detection task;an real-time citrus detection module was designed based on ARM architecture;a low-cost NPU computing chip is used to design a pluggable NPU cluster to replace the high-cost GPU for edge computing.After that,all these are applied to citrus detection task to reduce the cost of hardware,so as to realize the goal of low cost,low power consumption and high efficiency.And this module can be not only used independently,but also embedded in other systems as an auxiliary component,such as UAV,citrus picking robot,etc.The experimental results show that the AP of the algorithm model in this study reaches 94.16% on the citrus test set,and the average processing time of a single card on the graphics card RTX2080 is 24 ms.After the algorithm model is quantized and transplanted to the rk3399 + NPU cluster module,the AP is 93.98%.The average processing time of a single image on a single NPU is 83 Ms.after clustering four NPUs,the detection speed of the whole module is 47 images per second.