Scene Classification Model Design Based On Vision Sensor And Deep Neural Network

The processing ability of visual information has gradually become an important index to test the intelligence of the machine.The core of intelligence lies in the recognition and understanding of things.When the machine processes visual information,the main problem is how to understand the scene.It is one of the most important problems for human beings to understand the world and one of the key problems for machine vision.As an important branch of machine vision,scene recognition has been widely used in video monitoring,image retrieval,intelligent photography and other fields.The algorithm model proposed in this paper mainly adopts scene classification technology based on deep learning convolutional neural network technology.The main research contents are as follows:1.First of all for supervised learning research task,select based on the classification of "AI Challenger" 2017 scenarios challenge provided by the raw data as the foundation data,and carries on the statistical analysis of the data that includes 80 categories,each category of training data quantity is not the same,so I first of all,on the balance of tests and researches the algorithm of data processing,according to the actual experimental results,ultimately chose balance scheme based on the sampling was carried out on the types of data preprocessing,classification model for rare category recognition rate.2.Multiple classification models were designed for classification training and verification of the measured results.After a large number of experiments,the fusion model based on the weighted combination of three single models was finally determined to be used for classification.The three single models are a custom residues model,a network model with fine tuning based on Inception-v3,and a network model with fine tuning based on ResNet50.3.In the training process,data enhancement algorithm is used to randomly rotate,flip and cut the input data,which effectively improves the generalization ability andclassification accuracy of the model.At the same time,according to the theory of fusion model residual error model of training,has carried on the research of feature extraction algorithm,considering the complexity of the scene classification,this paper adopts Gabor filter technology global features of original image data are extracted,the global characteristics can better reflect the space layout,image filter unnecessary information,than the original image pixels have stronger power of expression.4.A large number of training experiments were carried out for model training,and various gradient descent(namely learning rate attenuation)schemes were designed to make the model as close to the optimal solution as possible in the end.Finally,the test set data of the scene classification challenge was tested in practice,and 80 scenes were automatically classified.The final top-3 accuracy rate was up to 98.20%,which was better than the current known test results on the verified data set.