Research On Pig Face Recognition Model Based On CNN

Traditionally,live identification of pigs generally uses the RFID technology,but the electronic ear tag and ear tag method will cause great pain to pigs,and the ear tag ear tag is easy to fall off during the pig’s activities,which increases the operating cost of the enterprise.Due to the high facial similarity of pigs and the influence of non-linear factors such as the environment of the farm,how to accurately identify the individual identity of pigs without harming the pigs,in recent years,deep learning technology in the field of computer image recognition has provided this problem a new solution.Convolutional neural networks in deep learning can automatically extract picture features from a large amount of data to complete the end-to-end image learning process.This article describes the research background and development status of pig face recognition,and summarizes the basic theories of CNN.By collecting a large amount of pig face image data and using Image Data Generator for data augmentation,a neural network model is trained under the deep learning network framework Keras.The classic network model LeNet-5 is improved,and the influence of the setting of training parameters on the accuracy of the model is studied.Aiming at the long training time and over-fitting phenomenon during the model training process,the idea of transfer learning finetuning is introduced.The main work is as follows:1.The pig face recognition dataset PFR-1 was constructed,and the augmented dataset PFR-2 was obtained through data augmentation technology.There are ten types of pig face pictures in the data set,with a total of 3,000 pictures,divided into a training set and a verification set at a ratio of 8: 2.The data enhancement through comparison experiments can significantly reduce the difference between the accuracy of the training set and the validation set of the model,and reduce the occurrence of overfitting to a certain extent.2.Construct a multilayer convolutional neural network.Based on the improvement of the LeNet-5 convolutional neural network model,the parameter settings of the model were determined through multiple sets of comparative experiments.The Maxpooling was used as the pooling layer,the Dropout layer was added,and the Softmax classification was used to finally determine the model’s convolution kernel size to be 3× 3,the activation function is Re LU,the Dropout ratio is 0.3,and the SGD optimizer.The accuracy of the model on the training set is97.63%,and the accuracy of the validation set is 88.50%.3.Use transfer learning techniques to train neural network models.Using the VggNet-16,Inception-V3,and ResNet-50 models trained on the ImageNet dataset,freezing the model convolutional layer through transfer learning finetuning,training only the fully connected layers,and comparing the VggNet-16’s accuracy is higher,it is more suitable for this dataset.4.The VggNet-16 model is improved,and the better AM-Softmax classification function is optimized based on the classification effect between classes.The neural network model with the lowest degree of overfitting when the penalty m is 0.3 is obtained.This model is on the validation set.The accuracy rate reaches 98.80%.Finally,the feature maps are used to visually show the different features from the bottom to the upper layers that are gradually extracted by the model in different convolutional layers,understand the feature processing process of convolution neural network in different convolution layers.