A lossless label target detection image processing method and device and storage medium

CN117115705BActive Publication Date: 2026-06-05CHENGDU KOALA URAN TECH CO LTD

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Patents(China)
Current Assignee / Owner: CHENGDU KOALA URAN TECH CO LTD
Filing Date: 2023-08-18
Publication Date: 2026-06-05

AI Technical Summary

Technical Problem

Existing image processing methods based on mixed data suffer from problems such as mislabeling, loss of object information, and class mutations in the field of object detection, leading to a decline in model performance.

Method used

By obtaining a rectangular area to be filled from the first image, erasing pixels and labels within the area, and then scaling the second image proportionally and pasting it into the area to be filled, combined with mosaic processing to remove cut bounding boxes that are prone to category mutations and shrink intersecting bounding boxes, the accuracy and integrity of the labels are ensured.

Benefits of technology

It effectively solves the problems of incorrect label assignment and loss of object information, improves the model's generalization ability and robustness to small targets, and enhances the performance of object detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN117115705B_ABST

Patent Text Reader

Abstract

The present application belongs to the technical field of machine vision target detection, and discloses a target detection image processing method, equipment and storage medium of non-loss label, comprising the following steps: obtaining a rectangular to-be-filled region in a first picture, and making the width and height of the to-be-filled region meet the requirements of preset values; erasing the pixel points in the to-be-filled region in the first picture, and clearing the label in the to-be-filled region; performing equal proportion scaling on a second picture and filling it to the same size as the to-be-filled region, and then pasting the processed second picture to the to-be-filled region. The present method effectively solves the problems of label error allocation, object information loss and category mutation existing in the existing data augmentation method based on mixed data, and simultaneously solves the problem of label damage during picture mixing processing. In the case of not increasing additional calculation cost, the processed picture with non-loss label can be obtained, thereby effectively improving the performance of the target detection training model.

Need to check novelty before this filing date? Find Prior Art