Compressor cutting feature recognition method and system based on 3D camera and deep learning

By using a 3D camera and deep learning approach, combined with multi-view active optical imaging and AI target detection, high-precision feature recognition and localization of waste household appliance compressors were achieved. This solved the accuracy and robustness issues of traditional methods in complex scenarios, and improved dismantling efficiency and material recycling benefits.

CN122244149APending Publication Date: 2026-06-19HEFEI SHANGJU IND EQUIP
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HEFEI SHANGJU IND EQUIP
Filing Date
2026-03-13
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing technologies struggle to achieve robust, complete, and precise feature recognition and positioning during the recycling and dismantling of used household appliance compressors. In particular, under complex industrial scenarios with strong reflection, weak texture, and structural self-occlusion, traditional 2D image processing and passive 3D scanning methods suffer from insufficient accuracy and detection blind spots.

Method used

Using a 3D camera and deep learning-based approach, high-precision point cloud data with normals is output through multi-view active optical imaging and AI target detection, combined with multi-level point cloud purification and fusion, for path planning of laser cutting robots.

Benefits of technology

It improved the success rate of feature recognition by more than 40%, the positioning accuracy by 75%, and the first-pass yield of laser cutting path from 62% to 99%, greatly improving dismantling efficiency and material recycling value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122244149A_ABST
    Figure CN122244149A_ABST
Patent Text Reader

Abstract

This invention provides a compressor cutting feature recognition method and system based on 3D camera and deep learning, including: (1) multi-view active imaging: a binocular structured light camera projects coded stripe light from multiple perspectives, and simultaneously acquires pixel-level aligned 2D texture map and depth map; (2) AI target detection: the texture map is input into the YOLOv8s model, and the bounding boxes of the tank top, bottom, and drain hole and their pixel center coordinates are output; (3) coordinate system transformation: the depth value d extracted from the depth map is combined with the projection inverse transformation formula to calculate the camera coordinates and map them to the robot base coordinates; (4) point cloud optimization: each feature point cloud cluster is separated; (5) parameter extraction: the outer normal vector of the tank top, the center of the bottom, and the axis of the drain hole are calculated; (6) multi-view point cloud fusion generates a complete feature model with normals. This invention integrates active optical imaging to suppress metal reflection defects, a multi-level point cloud purification chain to ensure sub-millimeter positioning accuracy, an edge feature retention rate of ≥98%, and a cutting path qualification rate of 99%.
Need to check novelty before this filing date? Find Prior Art