| The annual loss of stored grain in China is about 35 million tons,and the loss caused by insects account for 5%of the total loss.Most insects were first found on the surface of grain stack,so it is of great significance to know the occurrence of insects on the surface of grain stack for early control of insects.It is a common method for early detection of stored-grain insects to trap insects with sticky board.Therefore it is urgent to use intelligent means to solve the problem of detection and recognition the insects on the sticky board.Aiming at the automatic detection and recognition of six main stored grain insects,the following studies were carried out:1.The stored-grain insect image dataset for sticky board was established,this is the first relevant dataset in China.It includes object detection image dataset and paired image dataset.The object detection image dataset contains 20,827 images for training and testing of insect image detection network.The paired image dataset contains 8,134 images for training and testing of image classification networks.2.A stored-grain insect detection network based on SSD was proposed,which can detect and classify six kinds of insects trapped by sticky boards placed on the surface of grain heap.The detection tasks in this paper have the following characteristics:the similarity between different classes,the fixed position of cameras and trap,and the low correlation between classification and detection.According to the above characteristics,the following optimization of SSD was proposed:the obj ect bounding box regression strategy,loss function and feature extraction network structure.The test results showed that the improved SSD enhanced the effectiveness of insect detection,and the mAP reached 91.31%.3.Based on GANs,a lo,w-resolution image classification network for stored-grain insects was proposed in this paper.The average F1 value of image classification was 87.3%.4.Based on linear support vector machine,the fusion algorithm for obj ect detection network and low resolution image classification network was realized.The F1 value for fusion results is 95.22%,which is 12.95%higher than that of object detection network and 7.86%higher than that of classification network.5.In order to solve the problem of different operating system for client software and algorithm software and the shortage of resources for grain depot,based on the object detection network and low-resolution image classification network,the insects image recognition algorithm.service system and the insects image recognition client software were developed in this paper.At present,the algorithmic service system has been installed in 11 grain depots in China.By using this system,the warehouse keeper can discover the insects in warehouse as early as possible and control insect’s quantity stably according to the identification results.The system can reduce the fumigant dosage by 10%and save 130 man-hours per year on average.It improves the level of intelligent identification of grain depot insects in China. |
