Locusts Recognition Based On Deep Learning

Locusts plague,drought and waterlogging were once known as the three natural disasters,which have always been the focus of agricultural pest control in China.In order to help agriculture and forestry solve the problems of small locust targets,low recognition accuracy and delay in grasping the pest data under natural conditions,and reduce the loss of agricultural and forestry production,an improved YOLOv4-tiny locust target detection model is proposed based on deep learning,and a set of small locusts recognition system is designed on this basis.In order to be faster,the model uses CSPDarknet53-tiny as the backbone feature extraction network and Leaky Re LU as the activation function.Firstly,according to the characteristics of small locust targets,in order to improve the model’s ability of detail feature extraction and small target recognition under complex background,The Convolutional Block Attention Module(CBAM)is added to the input position of the enhanced feature extraction network and the output position of the first Convolutional result respectively.Then,aiming at the problem that there was no real-time data transmission for pest recognition in the past,the YOLO Head prediction module of the model was modified and embedded into the relevant cloud data platform to realize the real-time cloud data storage and analysis function of the model for locust infestation.Finally,the improved model was transplanted to Jetson Xavier NX embedded device,and a portable locust recognition system was designed by using Tensor RT to accelerate the inference of neural network.In order to verify the performance of the improved model and locust recognition system,experiments were carried out on self-made data sets and natural environments.Experimental results show that the performance of the improved model is improved in the self-made dataset,and the accuracy of the improved model is 1.69% higher than that of the original network.the average accuracy was increased by 1.57%,recall rate increased by 1.62%,the F1 value is increased by 2%,and the model size is only 23.6MB.With the help of the improved model and embedded equipment,the locusts real-time monitoring device is small and portable,and has strong locusts target recognition ability in natural environment,which can provide new ideas for agroforestry locusts recognition.