HUSKYLENS 2 AI Model Application: AI-Powered Industrial Component Inspection and Sorting Station
Project Overview
In industrial manufacturing and warehouse management, components with similar appearances often need to be identified, verified, and sorted before moving to the next stage of production. Traditional manual sorting relies on workers checking each item individually, which can be affected by experience, fatigue, and working conditions.
This project combines HUSKYLENS 2 AI Vision Sensor and UNIHIKER K10 to build a simplified industrial component inspection and sorting station. Users select a component from the unsorted area and place it in the recognition area. HUSKYLENS 2 runs a pre-trained AI model to identify the component, draw a bounding box around it, and display its name and ID on the screen. UNIHIKER K10 then displays the corresponding component information, allowing users to place the component in the correct sorting area based on the recognition result.
The entire process simulates a basic industrial workflow:
Component Input → AI Recognition → Information Verification → Sorting
This station currently recognizes three types of hardware:
- LIGHT SENSOR
- SOUND SENSOR
- LED MODULE
The system is not limited to these three components. By collecting new images, training a new model, and deploying it to HUSKYLENS 2, you can expand the system to recognize other components and explore additional product classification and automated inspection applications.
1 Getting to Know HUSKYLENS 2
HUSKYLENS 2 is an AI vision sensor with multiple built-in vision algorithms. It also supports deploying custom-trained models. It can capture images and perform AI inference independently, then send recognition results to a controller through its communication interface.
In this project, HUSKYLENS 2 is responsible for:
1. Capturing real-time images of the recognition area.
2. Running a custom-trained hardware recognition model.
3. Drawing bounding boxes around detected objects.
4. Outputting recognition results, including object names and IDs.
5. Sending recognition data to UNIHIKER K10.
2 Getting to Know UNIHIKER K10
UNIHIKER K10 is a compact learning board featuring a color display, camera, microphones, speaker, RGB LEDs, multiple sensors, and expansion interfaces. It supports graphical programming with Mind+.
In this project, K10 does not perform the hardware recognition itself. Instead, it serves as an information display and interaction terminal. Based on the results returned by HUSKYLENS 2, it displays the corresponding component name, image, and description.
3 Project Features
This project implements the following functions:
1. Recognizes three types of visually distinct hardware using a custom-trained AI model.
2. Draws bounding boxes around detected objects and displays their names and IDs on the HUSKYLENS 2 screen.
3. Sends recognition results to UNIHIKER K10.
4. Displays the corresponding component information on the K10 screen.
5. Shows an interactive welcome screen when no hardware is detected.
6. Guides users through manually sorting components based on AI recognition results.
7. Supports expanding the system to recognize more objects and handle additional scenarios by retraining the model.
4 Getting Started
4.1 Hardware Preparation
| Hardware/Material | Quantity | Function |
| HUSKYLENS 2 AI Vision Sensor | 1 | Captures images and runs the hardware recognition model |
| UNIHIKER K10 | 1 | Receives recognition results and displays component information |
| 4Pin Cable | 1 | Connects HUSKYLENS 2 to K10 |
| USB Cables | 2 | Used for program uploading, model deployment, or powering the devices |
| Light sensor | 1 or more | First recognition target |
| Sound sensor | 1 or more | Second recognition target |
| LED module | 1 or more | Third recognition target |
| Display board | 1 | Divides the station into unsorted, recognition, and sorting areas |
| HUSKYLENS 2 mounting bracket | 1 | Mounts the camera above the recognition area |
| K10 mounting fixture | 1 | Secures K10 in a position where users can easily view the screen |
| Power supply module or adapter | As needed | Provides stable, continuous power to the devices |
Tip: Keep the three recognition targets intact and make sure their main visual features are not blocked by labels, cables, or temporary fixtures.
4.2 Software Preparation
Prepare the following software and files:
- Mind+: Used to program UNIHIKER K10 and upload the program.
- HUSKYLENS 2 custom-trained model and installation package: Used to recognize the three types of hardware.
- Three component information images: Each image should be 240 × 320 pixels.
- One interactive welcome image: Displayed when no hardware is detected. The image should also be 240 × 320 pixels.
4.3 Preparing the Display Board
To make the user experience clear and intuitive, divide the display board into three main areas:
| Label on Display Board | Purpose |
| UNSORTED HARDWARE | Holds components waiting to be identified |
| RECOGNITION AREA | Holds the component currently being identified |
| SORTING AREA | Holds the three types of hardware according to recognition results |
Label the three sorting positions as follows:
- LIGHT SENSOR
- SOUND SENSOR
- LED MODULE
5 Training the Hardware Recognition Model
5.1 Define the Recognition Classes
This project uses three target classes:
| Class Name | Corresponding Hardware |
|---|---|
| LIGHT SENSOR | Light sensor |
| SOUND SENSOR | Sound sensor |
| LED MODULE | LED module |
Keep the class names consistent across the dataset, model outputs, program logic, and display assets. Inconsistent names may prevent the program from correctly matching recognition results.
5.2 Collect Training Images
Take photos of each of the three types of hardware. To improve recognition stability, collect images that cover a variety of conditions, including:
- Different placement angles
- Slight changes in object position
- Different distances and object sizes in the frame
- The actual background used on the display board
- Potential lighting changes at the exhibition venue
- An empty recognition area with no hardware present
- Distracting objects that look similar but do not belong to the target classes
- Make sure the hardware is clearly visible in every image. Avoid having fingers, cables, or reflections obscure important features.
5.3 Annotate Images and Train the Model
Upload the collected images to the model training platform.
Annotate the images using the following labels: LIGHT SENSOR, SOUND SENSOR, LED MODULE. Then start training the model.
Once training is complete, do not rely solely on the training results. Test the model using new images that were not included in the training dataset. Pay particular attention to the following:
- Can the model correctly distinguish all three types of hardware?
- Can it recognize components when they are away from the center of the frame?
- Does it produce false detections when the recognition area is empty?
- Does a hand entering the frame interfere with recognition?
- Do exhibition lighting conditions or reflections affect recognition?
If the model frequently misidentifies a particular component, collect more images of that component under different angles and lighting conditions, then retrain the model.
You can also upload test images through the File option to evaluate the model.
5.4 Export the Model Installation Package
After the model passes testing, export the model installation package compatible with HUSKYLENS 2.
Keep the original installation package intact. Do not modify its file structure or filenames.
6 Deploying the Model to HUSKYLENS 2
6.1 Connect the Device
Connect HUSKYLENS 2 to your computer using a USB cable.
Once the connection is successful, the corresponding device storage will appear on your computer.
6.2 Copy the Model Installation Package
Copy the exported model installation package into the designated model installation directory on HUSKYLENS 2.
6.3 Install the Model
On the HUSKYLENS 2 screen, open the model installation function, select Local Installation, and locate the model package you just copied.
Once the installation is complete, the new hardware recognition model will appear in the algorithm list.
Select Local Installation.
6.4 Check the Algorithm ID
Open the installed hardware recognition model and note its algorithm ID. You will need this ID in your Mind+ program to switch to the correct custom-trained model.
6.5 Test Recognition
Place each of the three types of hardware in the recognition area one at a time. Confirm that HUSKYLENS 2 draws a bounding box around each component and displays the correct name and ID.
If the model works well during computer testing but becomes unstable when installed on the display station, first check the camera height, recognition distance, background, lighting conditions, and component placement. Only then decide whether additional training data is needed.
7 Creating Display Assets for UNIHIKER K10
K10 displays different images depending on the current recognition status. This project requires four portrait images, each with a resolution of 240 × 320 pixels:
1. An interactive welcome screen displayed when no hardware is detected.
2. A LIGHT SENSOR information image.
3. A SOUND SENSOR information image.
4. An LED MODULE information image.
To maintain a consistent visual style, all four images use yellow, black, and white as their primary colors. Keep the layout consistent, including the title bar, product illustration, and information area.
Each component information image should include at least:
- Product name
- Product illustration
- Main functions
- Typical applications or key features
- Essential interface or specification information
Since the screen is relatively small, keep the text concise and prioritize the readability of the product name, illustration, and key information.
8 Hardware Connections and Assembly
8.1 Connect the Hardware
Use a 4-pin cable to connect HUSKYLENS 2 to UNIHIKER K10. Make sure both devices have a stable power supply.
After connecting the devices, run a desktop test first. Make sure K10 can successfully read the data returned by HUSKYLENS 2 before mounting everything onto the display board.
8.2 Mount HUSKYLENS 2
Mount HUSKYLENS 2 above the recognition area, with its camera pointing vertically downward.
The camera should cover the entire RECOGNITION AREA while minimizing the visibility of nearby sorting areas and users' hands.
After mounting, check the following:
- Is the camera securely mounted?
- Is the lens pointing straight down?
- Are all three types of hardware fully visible in the frame?
- Can the model reliably recognize objects near the edges of the recognition area?
- Can users easily view the bounding boxes and recognition results on the screen?
8.3 Mount UNIHIKER K10
Secure K10 next to the recognition area, with its screen facing the users.
Make sure the screen is not blocked by HUSKYLENS 2, its mounting bracket, or any hardware waiting to be sorted.
8.4 Arrange the Display Board
Place the unsorted hardware in the UNSORTED HARDWARE area and set up three clearly labeled sorting positions on the right.
Only one component should be placed in the central RECOGNITION AREA at a time to prevent multiple objects from appearing in the camera frame simultaneously.
9 Programming with Mind+
9.1 Add the Devices and Extensions
In Mind+, select UNIHIKER K10 and add the corresponding HUSKYLENS 2 library or extension.
9.2 Initialize HUSKYLENS 2
When the program starts, initialize HUSKYLENS 2 and switch to the custom-trained model using the algorithm ID recorded earlier. Then continuously read the camera's recognition results.
9.3 Display Component Information Based on Recognition Results
Read the name or class ID of the detected object, then use conditional statements to switch the content displayed on K10.
| HUSKYLENS 2 Recognition Result | Content Displayed on K10 |
| LIGHT SENSOR | Light sensor information image |
| SOUND SENSOR | Sound sensor information image |
| LED MODULE | LED module information image |
| No valid object detected | Interactive welcome screen |
When no hardware is detected, K10 displays the interactive welcome screen, prompting users to place a component in the recognition area.
9.4 Upload and Run the Program
Connect K10 to your computer, compile the program, and upload it to the board.
After uploading, test the following four states one by one:
1. The recognition area is empty.
2. Place a LIGHT SENSOR in the recognition area.
3. Place a SOUND SENSOR in the recognition area.
4. Place an LED MODULE in the recognition area.
Make sure the recognition results on HUSKYLENS 2 match the information displayed on K10.
10 User Interaction Workflow
Users can interact with the station by following these steps:
1. Select a component from the UNSORTED HARDWARE area.
2. Place the component by itself in the RECOGNITION AREA.
3. Observe the bounding box, name, and ID displayed on HUSKYLENS 2.
4. Check the component information displayed on the K10 screen.
5. Place the component in the corresponding SORTING AREA based on the recognition result.
6. Select another component and repeat the challenge.
To make the experience more intuitive, you can add a short instruction next to the display board:
Pick one. Scan it. Sort it!
11 Project Results
Once completed, HUSKYLENS 2 can recognize the three types of hardware and display their bounding boxes, names, and IDs on its screen. UNIHIKER K10 simultaneously displays the corresponding component information.
Users can then sort the components into the correct areas based on the AI recognition results, completing an interactive experience that simulates the process of visual inspection and sorting.
This project is an interactive prototype designed to demonstrate the basic workflow of AI visual recognition. Real-world industrial applications require more rigorous testing and validation of dataset size, recognition accuracy, processing speed, communication stability, safety mechanisms, and exception handling.
12 Project Extensions
12.1 Add More Hardware Categories
Collect and annotate images of additional components, then retrain the model to recognize more types of sensors, actuators, or electronic components.
12.2 Add Incorrect Sorting Alerts
Install sensors in the sorting areas to check whether users place components in the correct positions. Provide feedback through lights, sounds, or on-screen messages.
12.3 Add an Automatic Sorting Mechanism
Use servos, conveyor belts, or robotic arms to automatically move components into the corresponding areas based on AI recognition results.
This would upgrade the system from manual sorting to automated sorting.
12.4 Record Recognition Data
Record the number of detections for each component, recognition times, and abnormal results. Display statistics on the screen or generate reports on a computer.
12.5 Adapt the System to Other Applications
By changing the training data and display board layout, the system can also be adapted for:
- Industrial component recognition
- Warehouse material classification
- Product model verification
- Packaging content inspection
- Laboratory equipment management
- AI vision education and model training demonstrations
13 Conclusion
This project combines the custom-trained vision models of HUSKYLENS 2 with the programming and display capabilities of UNIHIKER K10 to create an interactive industrial component inspection and sorting station.
Throughout the project, we followed these steps:
Define Recognition Targets → Collect and Annotate Data → Train the Model → Deploy the Model → Connect the Hardware → Write the Program → Assemble the Station → Test and Optimize
Through this project, users can see how AI recognizes and distinguishes different hardware components while learning how a visual recognition model moves from training to practical application.
More importantly, the system is not limited to the three components featured in this demonstration. By retraining the model with data from different scenarios, HUSKYLENS 2 can gain new recognition capabilities and be adapted to a wider range of component recognition, product classification, and automated inspection tasks.
