Convolutional neural network prediction method based on rotation region

Abstract

The invention relates to a convolutional neural network prediction method based on a rotation region. The method comprises steps that step 1, a convolutional neural network characteristic graph and an area of interests based on a rotary rectangular frame are generated according to input images; step 2, a rotary area-of-interest pooling layer is utilized to acquire a characteristic mapping graph according to the convolutional neural network characteristic graph and the area of interests; step 3, the characteristic mapping graph is mapped to be a one-dimensional characteristic vector; step 4, the one-dimensional characteristic vector is classified to acquire a classification result; step 5, rotary rectangular frame regression prediction of the one-dimensional characteristic vector is carried out to acquire a regression prediction result; and step 6, a final prediction result is outputted according to the classification result and the regression prediction result. The method is advantaged in that a problem of positioning and identification difficulties caused by target rotation and an aggregation phenomenon of images is solved, and precise positioning is realized.

Description

technical field [0001] The invention belongs to the technical field of image analysis, and in particular relates to a convolutional neural network prediction method based on a rotation area. Background technique [0002] Due to the outstanding performance of deep learning in the field of image analysis, related research has been greatly promoted in recent years, especially in the field of image target detection and recognition, various classic frameworks have been launched one after another. [0003] Currently, the most classic framework models for detection and recognition include R-CNN framework, Fast RCNN framework and FasterRCNN framework. In the R-CNN framework, a large number of candidate rectangular frames are first obtained by a fast method, and then convolutional neural network features are extracted for each candidate frame area in the image, and then classified; due to the large number of candidate frames and the feature extraction, the efficiency is low . The F...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above problems in the prior art, that is, in order to solve the problem of difficult positioning and recognition caused by the rotation and aggregation of the target in the image, the present invention provides a convolutional neural network prediction method based on the rotation area, Include the following steps:

Images