Improved Convolutional Neural Network Algorithm And Its Application In Aquatic Plant Identification

In the development of botanical research and the application of agricultural technology,plant species and numbers are complicated and numerous,and identifying plant species is a daunting task.Due to the large number of aquatic plant species,it is important to construct a large-scale aquatic plant dataset,and aquatic plant species identification and classification is a very important task.According to many good properties and characteristics of convolutional neural networks,they can be used to recognize image features and distinguish common aquatic plant species.In this paper,training aquatic plants from different samples while extracting features to identify species has a very important role and far-reaching practical significance for aquatic plant identification.In this paper,a database of aquatic plants is constructed by crawling and manually screening images,and the following work is done:(1)Processing data.Building an aquatic plant database from pictures of 51 species of aquatic plants by crawler,and the whole images of aquatic plants were identified.Firstly,the images are manually screened to remove watermarked images and irrelevant images,then the images are cropped to the same size,and finally the images are enhanced with symmetric transformation,cropping,RGB channel conversion,rotation and other operations to expand the data,improve the sample size and mitigate the overfitting of the model.(2)In this paper,we propose an improved convolutional neural network model EF-Dense Net169 network,in which the activation function uses ELU activation function,the loss function uses Focal Loss function,the global average pooling method is used to optimize the network model,and finally the convolutional kernel size,the number of feature maps and related learning parameters are modified and adjusted.The existing Alex Net,VGGNet,Res Net,Dense Net and the improved model EF-Dense Net169 are used for comparison and analysis to extract the effective features of plant images,and the experimental results are obtained.The results show that the improved EF-Dense Net169 has the best recognition effect and the accuracy rate is significantly improved compared with other models,which can be applied in aquatic plant recognition detection.(3)Convolutional neural network and transfer learning fusion algorithm.Based on the convolutional neural network for recognizing aquatic plants to automatically learn to classify aquatic plant pictures,the migration learning algorithm is applied to migrate the parameter model of Image Net data to the aquatic plant detection model for recognition to achieve fast convergence in the multi-classification dataset of aquatic plants and avoid the problem of low accuracy of traditional image recognition.The results show that migration learning is suitable for classifying aquatic plant images and greatly improves the accuracy rate.