Research On Scene Recognition Algorithm Based On CNN And Multi-source Image

Infrared and visible image acquisition hardware different imaging principle is also different,image content information is also very different.With the development of society in recent years,object recognition and scene recognition have become the focus on research in computer vision fields such as unmanned driving and augmented reality.However,due to factors such as computer performance,GPU performance,and massive data,the resolution of visible light images tends to be low.In multi-source image applications,the recognition accuracy is often low,so image super-resolution is introduced.Utilizes super-resolution and blending techniques to obtain a comprehensive image with rich details,allowing viewers to accurately identify scenes.In the traditional visual field,the image reconstruction recognition process has the problems of long time-consuming algorithm,loss of detail information,low quality recovery,similar scene and low recognition efficiency.This paper combines the popular research methods at home and abroad to study these problems under the neural network model.The specific work of the paper is as follows:(1)The neural network(NN)theory and model are studied,and a super-resolution reconstruction algorithm based on low-quality visible light image is proposed.This method combines the sliding slice method to process the image and the simulated pyramid to obtain multi-scale image blocks.According to the feature point distribution,this paper proposes a mosaic convolution super-resolution reconstruction method for the first time.The experimental measurement indicators prove that the use of the details of the map can be used to obtain reliable details.(2)Studied multiple Neural Networks(NNs)models.Aiming at the problems of NNs model that only relies on the last layer of network,insufficient information of single source image,similarity of different contents in scene conditions and low accuracy of classification and recognition,this paper proposes a multi-source fusion neural network(FCNN)based on CNN to share parameters and reduce network complexity.In this paper,pixel fusion and convolutional neural network(CNN)model are used to learn multi-source images,and the fusion method and network model are improved respectively to learn rich comprehensive images.Compared with other methods,the experimental results show that,this identification method can quickly determine the picture category,improve the accuracy,and achieve the requirements of rapid identification.Compared with the performance of the verification algorithm,this paper carries out experimental verification and analysis on the set5,set14 and RGB-NIR data sets.By verifying the speed and accuracy requirements of the recognition algorithm,compared with the classical algorithm from the subjective and objective criteria and performance,the algorithm proposed in this paper is more effective and meets the accuracy requirements.