Multi-Object Recognition Technology For Street View Image

Because of the high-speed improvement of deep learning,computer vision turned into the fastest-growing technology.As a branch of object recognition,street view multi-object recognition is widely used in road supervision,driverless cars and other fields.Conventional vehicle detection performs feature extraction by direction gradient histogram and scale invariant feature transformation,and inputs the extracted features to the classifier.This method of manually extracting features as an image representation has inadequate abil-ity to adapt to complex scene changes,and the generalization ability is not good enough too.Recent years,object recognition algorithms based on convolutional neural networks(CNN)gradually show its unique advantages.Its advantage is that,the features are extracted auto-matically for data set,and has more invariance to changes such as deformation and lighting.Whether it is recognition speed or accuracy,there has been a good improvement.How-ever,at present,there is still a lack of object recognition in scenarios such as small objects recognition or with occlusion situations.Based on the above analysis and the application scenarios of actual street view object recog-nition,this thesis studies an object recognition algorithm based on the Mask R-CNN,by pre-procrssing the data base and impoving the algorithm,gets a better recognition accuracy.The main tasks are as follows:(1)For the actual scenario of this thesis,in order to obtain the most suitable object recognition algorithm,had studied the most representative object recognition algorithm models.Choose Mask R-CNN,as the object recognition algorithm model in this thesis,which has better processing effect in actual street view images with many small objects.(2)Based on the research,the network structure of feature extraction and the structure of the candidate window classifier are designed;and by reducing the number of network layers to improve the efficiency of the algorithm.The use of bilinear interpolation reduces the error in feature extraction of the region of interest and improves the accuracy.The histogram equalization and image sharpening are used to preprocess the data set,so that the accuracy of small object recognition in street view images is improved.(3)After preprocessing the data set using the above method,this thesis trains the improved algorithm and tests the actual effect,and compares the test effect with the YOLO algorithm.The tests show that the improved Mask R-CNN has good accuracy under different lighting and occlusion scenarios,and greatly improves the accuracy on small object recognition,and the accuracy is obviously better than the YOLO algorithm,which has reached the research goal of this thesis.