Research On Efficacy Balance And Flexibility Pest Identification Algorithm

Food security and crop growth have always been important issues of close concern to the people,and the frequent occurrence of insect pests is one of the important reasons for the decline in crop quality and yield.Recently,compared to traditional pest image recognition methods,deep learning-based methods in pest identification algorithms have been gaining momentum in recent years,with major breakthroughs in their performance.However,the improvement of recognition accuracy is often accompanied by an increasing number of network layers,and few researchers have paid attention to the phenomenon of high energy consumption of computer training with the increase of network layers,which limits the deployment and application of recognition algorithms on the mobile device side.In recent years,some methods by optimizing the network structure have been proposed,such as adder,binarization network,random node rejection and other algorithms.Unlike the traditional multiplicative convolution,the above algorithms occupy advantages in time complexity and space complexity,and require significantly lower energy consumption,however,they are also accompanied by a significant loss of accuracy.How to balance the relationship between accuracy and energy consumption and build a set of environment-saving algorithms with good recognition effect has become a challenging topic.To address this problem,this paper chooses additive convolution(Adder)as the theoretical basis and focuses on exploring the versatility of the generalized additive convolution(Adder)replacing multiplicative convolution(Conv)neural network structure optimization method,self-supervised learning recognition algorithm,and applying it to pest identification algorithms.The work in this paper is divided into the following aspects:a.A pest identification algorithm that can effectively balance identification performance and energy consumption is proposed.Specifically,firstly,we carefully analyze the characteristics of Adder and Conv and summarize their advantages and disadvantages;secondly,we analyze the structure of the backbone network(taking Resnet as an example)and try to explore the replacement ideas of different positions,different proportions and different modules;then,we determine an optimal additive-multiplicative convolutional hybrid network structure(AM-Resnet)through the experimental results under the premise of ensuring the difference in accuracy.AM-Resnet)with significantly lower energy consumption;finally,we conducted a specific and profound theoretical analysis based on different experimental results and migrated the principles to other backbone networks to verify the generalization and robustness of the network on more general data sets.b.A pest recognition algorithm that is both scalable and efficient is constructed.First,based on contrast learning,we design a self-supervised pest recognition algorithm SN to learn network features,and train a pest classification model with basic map recognition through unlabeled pest data;then,we introduce a downstream task,and through few epoches of model fine-tuning,we can achieve recognition accuracy comparable to the full-cycle training of supervised learning,bridging the accuracy gap of variants of AM-Resnets;finally,we propose a more efficient vision Transformer based on additive attention,which can effectively accelerate the convergence and training of the network.Finally,we propose a more efficient vision Transformer based on additive attention,which can effectively accelerate the convergence and training of the network.The number of output interfaces is adjusted by model fine-tuning to enhance the scalability of the model and facilitate fast iteration.