Research On Scene Recognition Based On Convolutional Neural Network And Object Detection

In recent years,with the gradual development of deep learning technology,its application scope in natural language processing,computer vision and other directions is gradually expanding,and the social demand is also increasing.Convolutional neural network in the field of deep learning has become the main method of image processing with its efficient feature extraction ability,which has great research value and commercial value.The traditional artificial scene image classification method has been unable to meet the current millions of image data.How to use the computer to analyze the scene image,and then realize the operation of classification,tagging,search and so on,has become a research hotspot of scholars.Although there are many research achievements in scene recognition,there are still many problems to be solved,such as the increase of parameters and calculation due to the deepening of the network,the low accuracy of the network with few layers,and the algorithm can not meet the recognition requirements.In view of the existing problems of scene recognition,this thesis studies a new convolution neural network object detection model,realizes the high-speed recognition of the scene image,and achieves a better balance in accuracy and speed.In the traditional Yolo v3 target detection model,the deep separable convolution method is applied to the original model network.Compared with the conventional Yolo v3 target detection algorithm,the parameter quantity of the new model is reduced by about 90%,and the speed of model training and object detection is increased by two times.This thesis verifies its performance in MIT indoor67 indoor scene image data set,which can achieve 62.4%accuracy,and can be applied to the environment with high real-time requirements,such as handheld devices.In this thesis,I find that the occurrence of various objects in the scene image is independent of each other.I calculate the prior probability of the occurrence of objects by object detection,then bring in the Bayesian formula to calculate the probability of the scene in the image,and finally take the maximum value as the result of scene recognition.By proposing that the occurrence events of objects are independent of each other,the scene recognition in this thesis conforms to the principle of naive Bayesian classification,which is convenient to calculate the posterior probability,thus reducing the amount of parameters and calculation of the model,and can achieve more than 90% classification accuracy.In this thesis,based on the above methods,a scene recognition system is completed.After experimental testing and comparing with the recognition rate of common scene recognition methods on the same data set,it is proved that the recognition accuracy of this model can reach a considerate level.