Research On Intelligent Breeding Method Of Chinese Soft-shelled Turtle Based On Deep Learning

At present,the breeding of superior species of Soft-shelled turtle is mainly carried out by traditional manual selection,but the disadvantages of traditional manual selection include slow speed,high labor cost and low efficiency.To solve this problem,this thesis proposes an intelligent breeding method of Soft-shelled turtle based on deep learning,which can finish the breeding work of fine species of Soft-shelled turtle more efficiently,accurately and conveniently.The main research contents include the following aspects:1.A large enough Chinese soft-shelled turtle dataset was constructed.The original images of the dataset were first collected manually by on-site shooting,and then expanded by numerical enhancement methods such as horizontal flipping,vertical flipping,brightness adjustment,and adding noise.Finally,de-mean normalization,data set size segmentation,graphic size unification and other operations are performed on the expanded data set.2.A feature-based intelligent selection model of Chinese soft-shelled turtle is built,and the intelligent selection of Chinese soft-shelled turtle is completed through several key features that can judge whether the Chinese soft-shelled turtle is standard or not.Complete training on the neural network.The experimental results show that the classification accuracy can reach 85.69% when using multiple feature combinations and using PCA for dimensionality reduction.3.A Chinese soft-shelled turtle intelligent breeding model based on model transfer learning was built,and the model parameters after pre-training were transferred to the Chinese soft-shelled turtle intelligent breeding model,and the training and testing were completed.The effects of different transfer learning models,different training strategies,and different optimization algorithms on the system classification effect are compared,and the classification effect is reflected in the accuracy rate and cross entropy loss value.The experimental results show that the Res Net-152 model using the Adam optimization algorithm can achieve an accuracy of about 92% and 90% on the training set and test set,respectively,by retraining the full layer of the pre-training network.4.An intelligent selection model of Chinese soft-shelled turtle based on shallow convolutional neural network was built on this model.Two different optimization strategies are compared,and the experimental results show that the classification effect is improved to some extent after adding the segmentation learning rate decay strategy on the basis of the original optimization algorithm.5.An intelligent breeding method for Chinese soft-shelled turtles based on the fusion of the improved models was proposed.First,the global average pooling was used to improve the network structure of each model.The experimental results showed that the accuracy and training efficiency of the improved models were improved accordingly.Then the improved model is fused based on model fusion,the fused model combines the respective classification results of the individual models,and the two model fusion strategies are compared.The experimental results show that the fused model is better than the single base model before fusion in terms of classification accuracy and system stability,and the final accuracy rate of the ensemble model using the average fusion strategy reaches 97%,and this thesis is completed.The goal of intelligent selection.