Zero-sample target detection model and establishment method thereof

Abstract

The invention discloses a zero-sample target detection model and an establishment method thereof, and belongs to the field of pattern recognition. The method specifically comprises: training an IoUGANaccording to visible RoI visual features, Gaussian random noise and a semantic embedding vector of a visible class; inputting the semantic embedding vector of the invisible class into an IoUGAN to obtain a visual feature of the invisible class; training a zero sample classifier by using the visual features of the invisible class; and combining the zero sample classifier with a feature extractor and a frame regression device to establish a zero sample target detection model. The IOUGAN is used for receiving the semantic embedding vector of the invisible class and generating a visual feature training zero sample classifier of the invisible class; the IOUGAN comprises a CFU, an FFU and a BFU; and, according to the zero-sample target detection model obtained by the invention, the position andthe category of the target can be accurately identified according to the input invisible test sample, and the practicability is relatively high.

Description

technical field [0001] The invention belongs to the field of pattern recognition, and more specifically relates to a zero-sample target detection model and its establishment method. Background technique [0002] Object detection methods based on deep learning have received extensive attention due to their excellent accuracy and real-time performance. However, the performance of detectors relies on large-scale detection datasets with fully annotated bounding boxes. The real world has a large number of categories, and it is often impractical to collect enough labeled data. The purpose of zero-shot target detection is to classify and locate new categories at the same time without training samples, which can avoid the above problems and does not need to collect a lot of labeled data. [0003] Zero-shot object detection can be performed in two spaces: the semantic embedding vector space and the visual feature space. Existing methods usually map visual features from predicted bo...

AI Technical Summary

Problems solved by technology

However, visual characteristics not only include intra-class differences, but also IoU (Intersection over Union) differences that are critical for target detection. Existing zero-shot target detection methods usually do not consider IoU differences, which may easily lead to accurate detection results of zero-shot target detection methods Problems of sex and usability

Images