Research On Animal Recognition Algorithm Based On Convolutional Neural Network

With the development of mobile devices,it is no longer a surprise to have a mobile phone or even multiple mobile phones.The application on mobile phones is also dazzling.Along with the high-speed update iterations of mobile devices,artificial intelligence is also rapidly developing at the same time.In order to develop a convenient mobile terminal animal identification application software,this paper uses convolutional neural network to study the animal image recognition on the mobile terminal.In this paper,the existing neural network structure is improved,some image processing algorithms are improved,and the animal images are trained by combining traditional image processing methods and convolutional neural network.Finally,the transplant is implemented on the mobile terminal neural network framework.The main achievements and innovations of this article include:(1)Study the data set balancing algorithm SMOTE algorithm,and propose an improved edge enhancement SMOTE algorithm.The animal is mainly located in the center of the image.In order to enhance the background diversity of the new sample of the data set,this paper proposes an edge-enhanced SMOTE algorithm with varying weights to increase the mixing of environmental information without affecting object mixing and using DenseNet121 convolutional nerves.The Stanford University dog data set was tested before and after the network was balanced.The recognition rate was 85.31%,which was nearly 10% higher than the literature.(2)In-depth study of current convolutional neural networks,proposes a new MobileNet-Beta network model based on MobileNet network improvement.Finally,the comparison of multiple network experiments shows that the DenseNet model is superior to the lightweight model such as MobileNet in recognition rate,but it is inferior to the MobileNet model in terms of recognition speed and model parameter size.In order to apply high-performance identification on mobile terminals,the MobileNet-Beta model was proposed,and the accuracy of the experiment on the AWA2 dataset was 2.86% higher than that of MobileNet and 1.27% higher than that of MobileNet V2.(3)This paper firstly verifies that SISURF has better performance than SIFT and SURF,and proposes a deep-learning training method that uses pre-training of neural network withmixed feature saliency maps.The map is pre-trained on the network,and then the original image is trained twice.Finally,from the experimental results,the pre-training method using the SISURF saliency map is shown to increase the recognition rate on the balanced AWA2 data set by 1.31%.In the balanced dog data set,the recognition rate increased by 0.88%.(4)Utilize NCNN,an open source framework for mobile terminals,to output the results of parallel network logic judgment and model migration on Android mobile terminals.In this paper,three trained models(MobileNet,MobileNet V2 and MobileNet-Beta network)loaded in the mobile terminal,and the final logical judgment of the model parallel output results,the final product test recognition rate of 95%.