Automated Recognition And Counting Technique For Agricultural Light-trap Pests Based On Deep Learning

China is a big agricultural country,and pests cause serious economic losses to agriculture each year.The premise of integrated control of pests and reduction of economic losses is to forecast the dynamics of pest populations in the field accurately and in real-time.Using light to monitor field pests is one of the main methods of measuring pests in China,it uses light to trap insects,and the next day,plant protection technician manually identify and count the pests that need to be forecasted.This method of artificially recognizing and counting pests relies on the subjective experience of plant protection technician,and there are problems such as heavy monitoring tasks,poor objectivity and non-real-time feedback.This thesis uses image processing and deep learning technology to study the insect images collected by the pests forecast excellent light placed in the rice field,and realize the automatic recognition and counting of rice pests.The main research contents and results include:(1)Research on background segmentation algorithm of agricultural light trap pest image based on difference image fusion.First,the B-channel image of the original image is edge-compensated.Secondly,the one-dimensional maximum entropy threshold segmentation algorithm is used to threshold the B channel of the original image.Then,the morphological operations are used to smooth the edges,remove the noise,fill the inner holes of the contour,and set the threshold to remove the invalid regions and the boundary connected domains.Finally,the obtained binary mask is mapped to the original image to remove the background.In view of the incompleteness of the binary mask extracted by the B-channel image alone,a new method is proposed to construct difference map using RGB three-channel image of the original image,to perform secondary segmentation and to fuse with the original binary mask,which ensures the integrity of the binary mask and lays a good foundation for subsequent segmentation and recognition.(2)Research on the segmentation algorithm of adhesion insects based on extremeerosion dilation.Firstly,By judging the properties of the adhesion image,the discriminant criterion for the adhesion area is designed.Secondly,using four different kernels to erode and dilate the adhesion image,and all connected domains are extracted to construct the candidate connected domain set.Then,a series of thresholds and discriminant criteria are set to filter the elements of the candidate connected domain set,and the repeated connected domains and the adhesion connected domains are removed.Finally,the remaining elements of the candidate connected domain set are extracted as the final set of individual pest connected domains.(3)Research on automatic recognition and counting algorithm of agricultural light trap pests based on deep learning.Firstly,preprocessing the individual insect image through rotation and size normalization.Then,the target area is judged to enter a convolutional neural network for small or large pest recognition for classification by area threshold.Finally,counting target pests that need to be forecasted.The test results show that resnet18 of the small insect classification models performs best,among which the recognition rate of WBPH and BPH is 90.7% and 85.5%,and the small non-target pests' recognition rate is 95.2%,resnet50 of the large insect classification models performs best,among which the recognition rate of PSB,RLF and SSB is 92.5%,95.0% and 95.2%,and the large non-target pests' recognition rate is 95.7%.This paper proposes a light-trap pest recognition algorithm based on image processing and deep learning to realize the recognition and counting of five kinds of pests in rice.The algorithm has high robustness and generalization performance,which can meet the needs of practical applications.