LearningCamCounts workpieces using images
Inspects products to show results

LearningCam is an Edge AI camera with a web app integrated into the camera itself. It counts workpieces passing through the counting line and separates good (OK) and bad (NG) items using AI models tailored to your workpieces. The camera processes the image internally, so there is no need to send images to a computer or cloud. You can view results and configure settings from a browser on tablets or multiple computers.

Old-style counters only know that something passed, but cannot distinguish between different workpiece models. Manual counting and visual inspection can tire inspectors and lead to mistakes. LearningCam uses images from the camera instead, and has 2 modes of operation as shown in the table below.

How does LearningCam differ from Task Cam?

LearningCam observes workpieces, while Task Cam observes the work process of employees at the workstation. If your challenge is counting or defect rate, start with LearningCam. If your challenge is whether employees complete tasks in order and how long each step takes, see Task Cam.

LearningCam starts up in 5 steps: mounting the camera, powering and networking, selecting models and drawing frames, testing before going live, and enabling AutoRun. All steps are done through the camera's web interface, no software installation required.

1. 1

   Mount the camera above the inspection point

   Position the camera to view the entire area where the workpiece passes. The Inspection Cam model specifies that a 12mm lens at 65cm sees an area of approximately 14x20cm, and at the maximum distance of 80cm, it sees approximately 16x23cm.

2. 2

   Power and network connection

   Supply power to the camera. The standard model connects via Wi-Fi 6, while the Pro model connects via LAN RJ45 to the Router or Switch. Tablets, computers, or screens on the same network can view the results.

3. 3

   Select models and draw frames

   Open the browser and go to the camera's web page. Select the AI model for the job, draw the detection area frame, then rotate or adjust the image resolution to clearly see the workpiece.

4. 4

   Try before using

   Press Test Run to view live results. If counting or separating is still incorrect, record real workpiece video as training data for the model and upload the new model.

5. 5

   Enable AutoRun and start the process

   Set the camera to log in and start self-inspection when powered on. Then bind Webhook or nimitFlow to receive results.

What does the camera's web page contain?

• Dashboard Status — Date and time, name of the currently used AI model, operating mode, count separated by Part No., and live camera feed with Start, Stop, and Reset buttons.
• Model Upload — Select or upload an AI model file from your device, and there is a button for synchronization to nimitFlow.
• Inference — Select Continuous (continuous detection) or Snapshot (single image detection) mode, set delay time before detection, draw detection area frame, rotate image, select image resolution, and press Test Run to preview results before actual use.
• Data Record — Take single images or record videos for up to 60 seconds, view recordings in reverse, download files individually or as a .zip package for model training, system warns when memory is full.
• Network — Set device name (Device ID), select Wi-Fi network in Standard model or connect LAN RJ45 in Pro model, view IP Address and MAC Address of the camera, and copy Webhook for sending results out. Counter Cam model selects automatic IP (DHCP) or static IP.
• System — Enable AutoRun to have the camera log in and start self-inspection when powered on, change password, reset to factory settings, and reboot.

The Inspection Cam model of LearningCam sends count data via Webhook (an automatic data transmission channel to the destination system). The message includes camera Part Number, processing time, OK count, NG count, trigger signal, and date-time. The factory's system or nimitFlow can receive this data for further reporting. You don't need to record counts manually.

• Webhook in the Inspection Cam model - sends data according to settings to the destination system
• nimitFlow - sends counting data into nimitFlow, an optional add-on
• TaskCam and Skeleton models - receive start, stop, count, and reset signals from IO pins or call through API to run alongside PLC or existing counters
• TaskCam and Skeleton models - record video and CSV table for each cycle, then upload to NAS via FTP automatically. If the upload fails, it will retry.

Processed images on the camera itself, no need to send to the cloud. Results and files are sent only to the destinations you specify.

LearningCam is suitable for jobs where workpieces pass through the same point repeatedly and require count data or OK/NG results from images. LearningCam is not suitable for jobs that measure size numerically for SPC (Statistical Process Control) and for checking employee work sequences. Use this table to check your job before scheduling a demo.

Size measurement tasks use InGauge with InQC. Work sequence verification uses Task Cam.

How to measure the performance of the camera?

Defect sorting should not be assessed by overall accuracy alone. For example, among 2,000 workpieces, 40 are defective. If a camera labels every piece as good, overall accuracy is 98 percent, yet it detects no defects. Assess two measures together: how many defective pieces are detected out of all defects, and how many good pieces are incorrectly rejected

The confidence score given by the model is not the probability of being correct until compared with real parts. We recommend running the camera alongside existing inspection before changing the process. Camera results are supplementary information. Your quality team remains the final decision-maker for releasing lots.

During B.E. 2568–2569, we installed 1 LearningCam Inspection Cam above a cap conveyor for a plastic water-bottle cap manufacturer. Its AI model was trained on real cap images from the factory to classify each cap as good (OK) or defective, such as excess flash (extra plastic along the mold seam) or incomplete molding, and count both categories. The Controller and Reject unit is designed to blow defective caps off the conveyor; rejection test results still need confirmation

An internal test report for January–March B.E. 2569 recorded inspection and counting of about 398 white caps per minute for 120 consecutive minutes, meeting the team’s criteria. These are inspection and counting results, not acceptance results for the complete rejection system. Blue caps tested with the white-cap model did not meet the criteria. Model adjustment and confirmation of full-system delivery are still in progress

During B.E. 2559–2560, we developed and installed a workpiece-counting system with a camera and a separate processing unit (not LearningCam) on a vibratory feeder for an industrial machinery customer. The system counts 5 types of workpieces by lane as they cross the counting line. The project has been delivered with an operating manual and work checklist

What we learned

• A model trained on one color cannot be directly applied to another color; it must be trained and tested separately according to color or production run.
• Lighting and camera angle affect the decision of the model; the camera must be fixed at a constant angle and tested under real lighting conditions on the production floor.
• Counting with a counting line requires the workpiece to run in a defined area; in the vibrating feeder case, workpieces that fall out of the lane will not be counted.
• Acceptance criteria must be agreed upon before testing, including conveyor speed, testing duration, and pass criteria, and recorded in the same report format for each round.

What does it mean to you?

Before requesting a quotation, we require images or clips of actual workpieces in all colors and models that need counting or inspection, along with examples of real defects found. We recommend starting training the model with the model you produce the most first, testing at the actual conveyor belt speed according to agreed criteria, then adding other models. If automatic defect removal is required, allocate a separate budget for the removal system and testing time apart from the camera.