Research On The Detection Of Plant Disease Based On Continual Learning

Detecting plant diseases has been a crucial area of study for worldwide sustainable development.Deep learning networks can be used to build automated methods for diagnosing plant diseases.Automated recognition networks based on multi-task learning have issues with properly allocating weights to each task,taking into account both the feature sharing component and the task feature representation,and having overfitting or underfitting issues during model training.This paper addresses the problems of parameter assignment,feature sharing and representation in existing multi-task continual learning models.The main work of this paper is as follows:Continual learning models add old features which were learned to new features to learn new concepts,or freeze parameters to mitigate forgetting.That may lead to slow degradation of features and a network could not assign weights to learning tasks well.To address these issues,a method is proposed to address catastrophic forgetting in order to achieve continual learning,incorporating the ideas of scalable feature learning and fine-tuned classifier learning,a two-stage training strategy to reduce neuron dependency and freezing in order to improve the continual learning methods.The process is divided into two stages that are performed sequentially.The first stage is scalable feature learning,which updates the existing feature representations and extends them with a new feature extractor trained on both new and previously memorized data.The second stage is classifier learning,which addresses class imbalance by using a balanced fine-tuning strategy.This method improves the plasticity of the model by regulating the direction of parameter optimization,and at the same time controls the stability of the model,and the consequences of catastrophic forgetting are alleviated.On the basis of scalable feature learning,a dynamic scalable feature method incorporating mask learning is proposed,combining Res Net residual structure and feature reuse of Dense Net,fusing dynamic scalable features to improve the model’s utilization of multidimensional feature information,and compensating for the drawback of continual learning not being able to take into account the feature sharing part and the feature representation of the task.The balance between the individual feature layers is improved by combining sparsity loss to better extract position information for the bottom layer of the leaf spot region.In order to improve the accuracy of plant disease classification,a two-stage training strategy for continual learning,a channel-level mask pruning to dynamically fuse feature representations,and a technique to reduce network parameters and solve the gradient disappearance problem are utilized.In order to achieve the impact of keeping useful features and deleting useless ones through sparsity,the loss function is finally tuned.The continual learning algorithm is enhanced in this work,and the deep learning algorithm is contrasted and analyzed with the upgraded algorithm.The results indicate that the accuracy rate has increased in comparison to the conventional algorithm,and both the backward and forward transfer rates have improved.This proves that the enhanced continual learning algorithm is more effective at detecting plant diseases and plays a small but significant role in preventing catastrophic oblivion.For sophisticated disease monitoring and forecasting in forestry and network implementation,this topic offers useful advice.