Recognition algorithm of street landscape in cold cities with high difference features based on improved neural network

Abstract

Abstract Aiming at the disadvantages of low recognition accuracy and high error rate when the current method recognizes features with high differences in urban street landscapes, the City Feature Recognition algorithm based on gray value and color feature preprocessing and convolutional neural network is proposed, and Object Recognition algorithm based on improved YOLO2 fully convolutional neural network. The urban street landscape high-disparity feature recognition solution uses two different algorithms after preprocessing the gray-scale image and the color image of the urban street landscape, respectively, in the data set of 1200 private photos and 5000 selected Mapillary Vistas. Experiments are conducted in the data set. In order to eliminate the influence of countries on the street landscape characteristics of cold cities, we selected pictures of cities adjacent to the same country in Experiment 1. In Experiment 2, pictures of non-cold cities were eliminated to reduce interference and make it more effective. Reflect the robustness of the method in this article. The experimental results show that in City Feature Recognition, our scheme can achieve an accuracy of 59%, which is a relatively high accuracy in experiments that only rely on pictures for prediction. In addition, in Object Recognition, the improved YOLO2 algorithm has a greater lead than the traditional RCNN algorithm in 8 categories of objects, and has achieved an accuracy of about 75% in categories with obvious characteristics.