The Lightweight Model For Insect Detection Based On Deep Learning

Accurate and timely insect detection is the key technology for monitoring and early warning of pest disasters in agriculture and forestry.Hitherto the forestry pest monitoring still has depended on human experts in great numbers to manually recognize and count pests.On the other hand,most of image-based insect detection methods depended on the lab environments controlled strictly or manual pre-processing which were not suitable for the actual field conditions.Considering the research status in the field of insect detection,this paper proposed the insect detection method based on the object detection to explore the feasibility of using deep learning technology for automatic insect detection in different scenarios.The lightweight detection models which could run on the embedded devices or the notebook computers were designed considering the application requirements fully.In this way,the detection process could get rid of the dependence on large workstations.On the one hand,the joint identification model was proposed which divided the detection process into two stages of location and recognition for the high-resolution insect data.The model processed the high-resolution insect images with low computational and storage overhead through location and recognition step-by-step.The experimental results showed that,the mean average precision of the joint identification model was 0.833 on the crop insect datase’t which was 13.5%better than the standard detection model.The single-picture average runtimes on the GPU server and the notebook computer were 45.17%and 66.65%faster than the high-resolution detection model,respectively.On the other hand,the lightweight detection model was proposed which was compressed for the computational cost and enhanced for the detection capability of the model through the optimization of the feature extractor,feature pyramid and prediction module.Then the proposed model was deployed on the.embedded device to verify the feasibility of the model running on embedded devices.The experimental results showed that,the average precision of the red turpentine beetle for the lightweight detection model was 0.746 on the bark beetle dataset.The single-picture average runtimes for the lightweight detection model on the Jetson TX2 and Raspberry Pi 3B were 0.448s and 23.44s,respectively.Through the experiments in this paper,the automatic insect detections in different scenarios were realized and the feasibility of object detection methods in the field of pest monitoring were effectively verified.The optimized lightweight detection model could provide the core identification technology for the early warning of the red turpentine beetle outbreaks.