Research On Recognition Algorithm Of Odonata Insects Based On Deep Learning

Computer recognition of insect images is an important application of artificial intelligence technology,which can free insect researchers from the boring and heavy task of species identification and greatly meet the urgent needs of different industries for insect species identification.Dragonfly is an ideal environmental indicator and natural enemy insect,and it is of great importance for environmental protection and pest control to realize accurate identification and localization of dragonfly insects in images with complex backgrounds.In addition,most existing insect image recognition systems use traditional pattern classification techniques,which mainly rely on manual segmentation of target regions and have limited classification accuracy.Therefore,this paper explores the deep learning-based insect image recognition technology for dragonflies and builds a practical recognition system based on it.The main work and research results include the following four aspects.(1)A dataset of 8871 images of 206 species of Chinese dragonflies was compiled and labeled,and a deep neural network-based target detection technique was implemented for target localization of dragonflies.By experimentally evaluating several mainstream deep target detection algorithms,the most suitable model for this problem was selected and improved to achieve better performance on dragonflies,while the number of candidate frames was reduced to improve the real-time performance of the algorithm,the anchor frame size was improved to improve the detection accuracy by introducing a clustering algorithm,and finally the model was compressed to further improve its portability.(2)The attention-based insect recognition algorithm is implemented for 46 types of dragonflies.The network model that works well in the dragonfly recognition task is firstly selected as the base network,and the recognition performance is effectively improved by proposing a random edge pixel discard mechanism,designing a region suggestion network and improving the loss function.(3)Combining some modules of hash learning image retrieval algorithm and improved dimensionality reduction algorithm,a discrimination algorithm for unknown dragonfly species is proposed.This algorithm enables the system to identify dragonflies in the input image that are not included in the identifiable species,and add the identified images to a new class pool in order to accumulate a data base for subsequent research.(4)A basic insect recognition system for dragonflies has been built.The system automatically detects the dragonfly targets in the user-input dragonfly images and feeds them into the recognition model for recognition,and then returns the recognition results in real time.In addition,the system will automatically store the dragonfly images identified as unknown species in the database of unknown species,so as to accumulate data for subsequent research.