Fine-grained image detection method and system based on improved RA-CNN

Abstract

The invention discloses a fine-grained image detection method based on an improved RA-CNN. The method comprises the following steps: S1, preprocessing a training image to obtain an image vector code and a category vector code of the training image; S2, training weak supervision by using an improved RA-CNN model according to the image vector code and the category vector code of the training image to obtain predicted bounding box information; S3, using a training picture marked with a bounding box as input, comparing the bounding box predicted in the step S2 with the marked bounding box, calculating a loss function, and performing strong supervision training to obtain a trained image detection model; and S4, performing gray processing and vector normalization on a to-be-detected image to obtain an image vector code of the to-be-detected image, and inputting the image vector code of the to-be-detected image into the trained image detection model to obtain an object category and bounding box information in the to-be-detected image. The invention further provides a corresponding fine-grained image detection system based on the improved RA-CNN.

Description

technical field [0001] The invention belongs to the technical field of target detection, and more specifically relates to a fine-grained image detection method and system based on improved RA-CNN. Background technique [0002] Since the convolutional neural network has emerged in computer vision, the research on deep learning has become more and more popular, and algorithms have emerged in an endless stream. Regarding the classification and positioning of fine-grained image objects, before the emergence of convolutional neural networks, most of them need to rely on a large number of manual annotations to mark the position of objects in the image and accurate local information, and then perform feature construction on highly distinguishable regions. The model is then classified with a classifier. The representative is a feature encoding method based on local regions proposed by Berg et al., which can automatically discover the most discriminative information. After the emer...

AI Technical Summary

Problems solved by technology

After the emergence of convolutional neural networks, research based on strong supervision has risen on a large scale. R-CNN (Region-CNN, regional convolutional neural network) uses selective search to avoid violent enumeration of candidate regions, but since each frame must After one pass of classification, there are many repeated calculations of feature maps

Images