Neural network training method, object search method, device and electronic equipment

Abstract

A neural network training method, object search method, device and electronic equipment are disclosed. The training method of the neural network includes: obtaining an initial feature map from an input image through a first neural network; inputting the initial feature map into a second neural network for generating object candidate frames to obtain a shared feature map and a first loss function value; The shared feature map is input into a third neural network for object detection, a fourth neural network for object re-identification, and a fifth neural network for instance segmentation to obtain second to fourth loss function values, and based on the first One through fourth loss function values ​​update the first through fifth neural networks. In this way, it is helpful to extract features sensitive to attributes, so as to perceive more comprehensive attributes of objects and improve the accuracy of object search.

Description

technical field [0001] The present application relates to the field of deep learning, and more specifically, relates to a neural network training method, object search method, device and electronic equipment. Background technique [0002] In the field of deep learning, the pedestrian search problem refers to searching for the target pedestrian with the same identity as the pedestrian to be queried from the pictures taken by different cameras under the non-overlapping view. The whole search process can be regarded as a unified realization of pedestrian detection and pedestrian re-identification. Among them, pedestrian detection refers to the use of computer vision technology to determine whether there are pedestrians in an image or video sequence and give precise positioning. Pedestrian re-identification is to process the set of pedestrians to be queried that has been cut into a fixed size, that is, to request the set of pedestrians to be queried with a fixed-size pedestrian...

AI Technical Summary

Problems solved by technology

Different from the person re-identification problem, the data processed by person search are images containing complex original scenes, that is, the size and position of the target person are uncertain

Therefore, whether the detection result of the pedestrian is accurate or not, and the distance from the camera will affect the accuracy of the search.

Images