Research On Diptera Insect Classification Technology For Mobile Platform Based On Neural Architecture Search

With the increasing frequency of international economic exchanges and the convenience of transportation,geographical barriers are gradually being eliminated.As a result,there is an increase in the exchange and penetration of species around the world.Marine transportation is the most important mode of transportation in international trade,accounting for more than two-thirds of the total transportation volume.With the frequent exchange of foreign trade,port transportation has developed rapidly,but the invasion of alien species in ports has become increasingly serious.Diptera,a common invasive alien species,are particularly adaptable to new environments and often destroy the local ecological environment by breeding in large numbers.This endangers the survival of local endangered plants and animals,while also causing a loss of biodiversity.Therefore,it is of great significance to strengthen the classification and identification of Diptera to protect biodiversity and ecological security.In this thesis,neural architecture search is utilized as the fundamental network model.Based on different application scenarios,a recognition and classification method of Diptera insects is proposed using differentiable neural architecture search.(1)In high recognition accuracy scenarios,the problem of low recognition accuracy caused by similar characteristics of Diptera insects was addressed by proposing a neural architecture search method based on a multi-feature search cell.By designing a multi-feature search cell,the output features of the previous layer were added to each layer search cell,enabling each layer search cell to contain more target features.This enhanced the richness of the target features and the stability of the neural network.An attention module was incorporated into the search space to improve the feature recognition ability of Diptera insects.Finally,mixed precision quantization was introduced during the training phase to reduce the computational resource overhead.(2)In low latency scenarios,the network model is plagued by high latency and a large number of parameters.This thesis proposed a lightweight neural architecture search method.Firstly,the lightweight search space was redesigned,with the introduction of Ghost modules and stochastic pooling.The network model parameters were reduced through lightweight candidate operations.Partial channel sampling was used for model search,and the center loss function was used to improve the edge normalization process of the model.The number of sampling channels were reduced without sacrificing accuracy,which improved the search efficiency and greatly reduced computational overhead in the search process.We ported the proposed method to the NVIDIA embedded mobile development platform and tested the network model performance in a real environment.The experimental results show that the classification accuracy of Diptera insects based on the multi-feature search cell reaches 98.13%,the model parameter amount is 12.1MB,and the latency is 8.4ms.The classification accuracy of Diptera insects based on the lightweight method is 97.51%,the model parameter amount is 10.1MB,and the latency is 6.98 ms.The proposed method surpasses existing methods of Diptera insect classification and meets the needs of identifying invasive species on mobile platforms.It effectively improves the ability to identify alien organisms and reduces the invasion damage caused by invasive species.