A Real-Time System For Detection Of Oilseed Rape Pests Based On Deep Learning

Oilseed rape is one of the most important oil crops in China.It is vulnerable to pests during the growth process,therefore the yield and quality of oilseed rape are easily affacted to varying degrees.Based on this,accurate detection and rapid diagnosis of pests are critical to improving the pertinence of pesticide application decisions and the accuracy of crop yield prediction.Traditional pest identification relies on artificial identification.The method is time-consuming and laborious,and has the problems of subjectivity and hysteresis,which cannot meet the growing demand for pest detection.Therefore,it is of great practical value to seek an objective,efficient and rapid pest detection method to solve problems.In order to realize the real-time detection of oilseed rape pests,this paper mainly carried out the following work:(1)By collecting images of oilseed rape pest from simple and complex backgrounds,a typical rape pest dataset containing 3,022 images of 12 categories was created,which was named RPDS.Among them,a pest image acquisition system was designed to collect pest images in the laboratory environment(simple background).Finally,the training images were increased from 2,115 to 10,575 by means of data amplification.At present,there is a lack of publicly available pest data set in the agricultural field.The RPDS data set can not only be used for the research in this paper,but also has certain universality,which can be applied to other insect pest identification and pest detection work.(2)At present,the existing object detection algorithms cannot balance the contradiction between detection speed and accuracy.In addition,oilseed rape pest dataset has problems such as the small number of samples,small size of targets,diverse posture and easy to be covered.In order to solve the above problems,this paper improved the SSD object detection method and proposed a new method called F-SSD-IV3,which was suitable for pest detection of oilseed rape.(1)Inception V3 was used to replace the VGG-16 of SSD algorithm as the basic network to improve the detection accuracy of model,the mAP of the SSD object detection method was raised from 0.6411 to 0.6812.(2)A feature fused method was designed to fuse context information,and the feature maps with different scales were fused by cascading modules.This method solved the problem that SSD was difficult to detect small targets,and the mAP was increased to 0.7417.(3)This paper used strategy that combined Soft NMS with Softer NMS,which improved the inadequacy of the NMS strategy of the original SSD algorithm when handling object overlap.After experimental comparison,the missed detection rate was reduced from 9.44% to 2.31%,and obtained more accurate results by further adjusting the output rectangular coordinates.The results showed that the mAP of the F-SSD-IV3 object detection method designed in this paper was 0.7481,and the detection speed of single image was 0.076 seconds,realizing high-precision detection of typical pests in oilseed rape.(3)To overcome the problems caused by insufficient training data and unbalanced number of samples between classes,data augmentation and increasing Dropout layer were used to reduce over-fitting and improve the generalization ability of the model based on the oilseed rape pest detection method called F-SSD-IV3.(1)After data augmentation,the brightness,contrast and saturation of the image were randomly adjusted,and the image was flipped,rotated,cropped and shifted.The mAP value was increased to 0.8105.(2)The ratio of samples number between the classes was controlled.When the ratio was 1:1.5,the best detection performance was obtained,and the mAP value reached 0.8204.(3)The Dropout layer was added and different probability p was set to randomly suppress some neurons in the hidden layer.When p was 0.8,the mAP value was as high as 0.8417.The detection performance of the model was further improved by these optimization methods.(4)This paper made a comprehensive summary and research on the object detection technology and related methods based on deep learning,and provided a solid theoretical foundation for the new target detection method.In this paper,the proposed method F-SSD-IV3 and several classical target detection methods(SSD300,Faster R-CNN and R-FCN)were compared on the RPDS to analyze the advantages and disadvantages of each method.The results showed:(1)Among them,the SSD300 had the fastest speed,and the speed of detecting a single image was 0.048 seconds,but the detection accuracy was only 0.6411.(2)The detection accuracy of Faster R-CNN and R-FCN was less than 0.68,and the speed of detecting a single image was about 0.15 seconds.(3)The F-SSD-IV3 had the highest detection accuracy,and the mAP was as high as 0.7481.The single image could be detected within 0.076,achieving a good balance of accuracy and speed.(5)In this paper,the trained F-SSD-IV3 model was deployed to the application system of mobile devices,and a real-time pest detection system based on Android platform was implemented.It basically meeted the function of pest detection,including four modules: image acquisition module,image preprocessing module,rape pest detection module and result display module.By testing 303 images,the results showed that the detection results were basically returned within 1 seconds,and the detection accuracy was as high as 89.77%,the missed detection rate and false detection rate were 7.92% and 2.31%.The results showed that the system had certain practicability and could accurately detect oilseed rape pests in field environment.