Scene Image Recognition Method Based On Convolutional Neural Network With Saliency Features

Unmanned motion platforms have been widely used in the fields of national defense,medical care,industry and agriculture due to their advantages in autonomous reasoning,planning,and control during navigation.How to complete the autonomous and intelligent navigation process is an important part of the unmanned motion platform to realize its system functions.Without relying on satellites,because traditional inertial,radio and other navigation methods have low long-endurance accuracy and low degree of autonomy,researchers focus on brain-like navigation technology,and image recognition algorithms are used as location nodes in brain-like navigation.The key technology of error correction has high requirements on the recognition accuracy of the algorithm.In order to achieve high-precision scene image recognition,this paper combines its saliency detection algorithm based on the optimization of the convolutional neural network strategy,and proves the feasibility of the proposed method through a series of experiments.The related work is as follows:First,a systematic description of image recognition algorithms is given,and the characteristics of different recognition algorithms are introduced.A recognition scheme based on convolutional neural network is proposed,which uses Tensor Flow to build a network framework,completes programming in the Py Charm environment,and completes hardware implementation based on the Xavier module.In addition,simulation experiments are performed on the NUC dataset,which basically proves the feasibility of the algorithm.Secondly,the L2 regularization method is used to optimize the structure of the basic network framework,which strengthens the stability and the generalization ability of the algorithm in complex environments,and improves the accuracy to a certain extent.Then,the saliency detection algorithm and the data enhancement algorithm are used to improve the scene image recognition algorithm,which reduces the interference of the background information in the image in the feature learning process,and at the same time enables the model to learn more abundant effective features.Experimental results show that the recognition accuracy rate on the NUC dataset has increased from 84.5% to 95.3%.Finally,in order to prove the general applicability of the method,this paper experiments on the NUC dataset,the Nord Land dataset and the Garden Point dataset,and the recognition accuracy rates were all above 90%.Meanwhile,this paper has also proved the real-time performance of the method.The recognition speed of image prediction on the NUC dataset is0.02464s/frame,which means that the algorithm in this paper has reached the expected requirements in terms of recognition accuracy and real-time performance.