中文
LINGBANG · VisionAgent Hands-on

Rubber & Plastic Products
Defect Inspection Field Story

Shrinkage, short shot, and flash on black molded parts—from few-shot semantic segmentation labeling, to tensor flow, event orchestration, and state gating—turn appearance defects into online rules you can detect, count, and block.

Project 0811-66 Defects: shrinkage / short shot / flash Semantic seg · few-shot train Edge runtime monitoring

Defects on rubber and plastic moldings are often fine and scattered: local short shots, sink marks, flash. The hard part is not capturing a clear photo—it is marking defect regions steadily at line takt, then wiring the result to release or reject.

This article follows a field result video and real VisionAgent pages: label & train → workflow → event orchestration → state machine → parameters → live demo. Screenshots match the product UI; content aligns with the black molding and NG labels in the video.

E0 · NG
Shrinkage
Local sink or collapse from uneven wall thickness / cooling—hurts sealing and mating faces.
E1 · NG1
Short shot
Underfill / missing material; segmentation masks the short-shot zone—often the main defect.
E2 · NG3
Flash
Flash, overflow, extra material; area thresholds cut false positives and overkill.

Spot visual checks miss fine defects

Short shots and shrinkage are easy to miss by eye. Customer returns cost far more than watching one more frame on the line. You need reproducible masks and NG stats.

01

Semantic segmentation: label & train

SEG

Side-by-side labeling teaches the model with few samples

Segmentation · train & infer

Open Semantic Segmentation: original on the left, result on the right; pick defect type NG1 Short shot. Use polygon / brush tools to outline the short-shot area; filter Annotated / Unlabeled and Good / Defective on the right. Libraries can reach hundreds of images. After saving labels, run migrate / normal inference / accelerated inference / detect.

  • Defect dictionary aligns with page params: short shot, leak, flash (add shrinkage on the line if needed);
  • Few samples start training; for a new part number, add a few labels and fine-tune—no algorithm rewrite.
Semantic segmentation labeling
Screenshot loading

▲ Semantic segmentation: dual-pane labeling · NG1 short shot

Labeling is the train entry—process sees color masks; quality owns the classes.
02

Workflow: five-operator tensor flow

FLOW

Drag-and-drop the inspection chain

Operator library · No-code

Create project “0811-66”, then drag and connect from the left library:

Industrial Camera→ Image Preprocess→ Semantic Segmentation→ Event Prediction→ State Machine
Industrial Camera
Overhead full outline of the molding, covering holes and flash edges; multi-channel / RTSP supported.
Image Preprocess
Denoise, enhance, normalize—stabilize black speculars vs beige trim contrast.
Semantic Segmentation
Output masks for short shot / leak / flash; sample efficiency and speed per product ratings.
Event Prediction
Watch segmentation signal changes; raise shrinkage / short shot / flash events.
State Machine
FSM / PLC scheduling: good release or NG block, driving status-monitor checks.
Tensor-flow canvas
Screenshot loading

▲ Tensor flow: Industrial Camera → Preprocess → Segmentation → Event Prediction → State Machine

The graph is the program—change defect policy without waiting on developers.
03

Event orchestration: GPU Event Orchestrator

EVENT

Turn defects into subscribe-able signals

Primitives · FSM events

Open the GPU Event Orchestrator. The right-side state list shows three station events (same as the runtime board):

  • E0 Shrinkage, E1 Short shot, E2 Flash—one primitive each;
  • Drag Object Exists / Does Not Exist, count / width, etc. to build decision logic;
  • Select an event on the right first, then configure detection on the canvas; save for the state machine to subscribe.
GPU Event Orchestrator
Screenshot loading

▲ GPU Event Orchestrator: E0 Shrinkage · E1 Short shot · E2 Flash

Detections become events that runtime checks and NG stats consume directly.
04

State machine: pending → good / NG

FSM

State chart locks release gates

Event tensor
  • Root S0 Pending; shrinkage / short shot / flash each enter their NG state;
  • All related events pass → good release; any NG event takes the block branch;
  • Same source as the runtime “status monitor” triple check—shrinkage and short shot may pass first while flash is still pending.
State Chart Editor
Screenshot loading

▲ State chart: pending splits to shrinkage / short shot / flash NG or good

Defect policy is enforced by the chart—not by operator habit alone.
05

Camera & parameter setup

PARAMS

Page params · defect dictionary & display

Parameter Config

Open Parameter Configuration → Page Parameters. Left-side defect stats align with label classes:

  • Example rows: NG1 Short shot, NG2 Leak, NG3 Flash with counts; add / delete defect types;
  • Result display: show mask on, show defect name on, font 40, line width 2—matching the live overlay;
  • Model params page sets crop, epochs, confidence and area thresholds (e.g. short-shot confidence 0.65, area 500);
  • On the Industrial Camera node, set channel, exposure, and ROI so the full molding is framed.
Parameter configuration
Screenshot loading

▲ Parameter Config · Page params: defect types and display settings

NG dictionary, thresholds, and overlay style are shared by the board and reports.
06

Live demo: how to read the result

RUN

Image, events, and verdict on one screen

Edge runtime · result view

After the chain is configured, go live and watch the result. On the left is the industrial-camera feed: a black shaped molding on a light table, beige trim and holes clear, pink and orange patches marking what the system is watching. On the right, check shrinkage, short shot, and flash one by one—a green check means that item passed; the header shows running, and below is the Detection status.

  • Left: live scene and defect paint—does it match the part?
  • Right: three events check off in order—progress at a glance;
  • Verdict: all pass → release; any fail → NG and log the matching code.
Runtime monitoring
Frame loading

▲ Live view on the left, event checks on the right

Detection frame
Frame loading

▲ Running: masks mark regions of interest

▲ Field video: rubber & plastic defect inspection result clip

Config, recognition, event checks, and NG verdict share one UI—ready for same-day acceptance.
07

Comparison & takeaways

AreaTraditionalVisionAgent rubber/plastic
See defectsHuman spot checksSemantic seg marks short shot / shrinkage / flash live
Set standardVerbal rules, paper samplesPage-param NG dictionary + confidence / area thresholds
ReleaseOperator habitEvent orchestration + state-machine gates
New part #Wait for algorithm retrainFew-shot label / migrate
RecordsPaper NG logsAutomatic defect counts and process trail

Four questions run through the project: What does the system look at, what counts as a defect, when to block, and how results enter stats.

Closing

Rubber/plastic inspection is not about spotting one short shot—it is chaining label & train → segment → raise event → state gate → NG stats into a flow the floor will use and quality can audit.

In one line: segmentation finds defects, event orchestration reports types, the state machine gates release, and the params page owns the NG dictionary and display—no-code turns a spot-check station into online defect judgment.
I
Ivan
Industrial Vision Solutions Expert · Author

Master’s degree in automation and machine vision. Over a decade in industrial vision—from classic AOI algorithms to deploying vision foundation models on the factory floor. Delivered 30+ inspection and process-monitoring projects across 3C electronics, appliances, automotive parts, rubber & plastics, and new energy.