Research On Small Target Detection Method In Insect Growth Stage Based On Convolutional Neural Network

Insect growth period detection is a branch of insect detection.It is one of the key technologies in biological research,insect quantity protection and identification.The traditional insect growth period detection algorithm can not meet the requirements of realtime and accuracy.Therefore,this paper uses deep learning technology to detect the growth stage of insects.Among them,YOLOv3 algorithm regards target detection as a regression problem to detect the whole image,which improves the detection speed and morphological changes of large targets in the insect growth stage,but there are missed detection of small targets in the image.In order to solve this problem,the network structure,loss function and anchor box of YOLOv3 network model are adjusted to improve the positioning and detection ability of small targets.The main contents of this paper are as follows:(1)Self made data set.In this paper,the common butterflies in nature as the research object.Using crawler technology to collect butterfly growth stage pictures,using data enhancement technology to expand,annotate butterfly growth stage pictures.From egg,larva,pupa to butterfly,it forms its own data set and takes it as training sample,test sample and verification sample.(2)Adjustment of network structure.The original network structure of YOLOv3 has complex convolution layers,and the features of small targets extracted by multi-layer convolution are not obvious,which is easy to cause small target missed detection.In order to improve this problem,this paper proposes to take Res Net34 as the backbone network of YOLOv3,improve the residual network structure,reduce down sampling and enhance the feature extraction performance of small targets.(3)The improvement of loss function.The original loss function is based on the open root difference of width and height,which does not take into account the size difference of the detected small target in the sample image due to the different distance.In order to reduce the influence of the large difference between the target frames,MS-SSIM loss function is introduced to replace the original frame position loss function,and the overlap ratio,center distance and aspect ratio between the candidate frame and the real frame are fully considered.(4)Anchor boxes update.YOLOv3 uses K-means method to cluster VOC data set.Considering that the K-means method is too sensitive to data initialization,and this paper is a self-made data set,so the K-means + + clustering method is selected to re select the K value,and anchor boxes suitable for self-made data set is designed.Based on the above improvements,experiments are carried out on self-made data sets.The experimental results show that the improved YOLOv3 can quickly identify the growth stage of insects compared with other target detection algorithms and small target detection algorithms,which verifies the timeliness and accuracy of the improved YOLOv3 based on convolution neural network.Finally,the insect growth stage recognition system is designed,put into real life,and an convolutional neural network system to improve YOLOv3 is realized.