Reverse Vending

Building a Real-Time Packaging Recognition System with Edge-Optimized AI

Edge Devices Computer Vision Data Engineering

With sub-500ms detection and lightweight packaging, the computer vision model is built for frictionless scaling, ready to support hundreds of machines in the national deposit return network. Challenge Accepted!

Building a Real-Time Packaging Recognition System with Edge-Optimized AI

Process & Story

Our client builds intelligent machines that automate the collection of recyclable bottles and cans, known as Reverse Vending Machines. To ensure the accuracy of returns and prevent fraud, they needed a solution that could verify an object’s shape, material, and barcode, including distinguishing eligible packages from lookalikes and damaged containers.

We designed a lightweight machine learning system tailored to their hardware and performance constraints. The result? A scalable, fast packaging recognition system ready for nationwide deployment.

Problem

The client’s machines operate in the context of a deposit return scheme. That means they don’t just need to accept recyclables, they must:

All of this needs to happen in real time on a compact Raspberry Pi 5, inside a public-facing machine.

Solution

We designed and delivered a real-time object recognition system that helps our client verify recyclables and prevent fraud in deposit return machines. The solution combined custom dataset development, lightweight model training, and edge deployment—all optimized to run on Raspberry Pi 5 within a strict 500ms time budget.

Outcomes

Custom dataset and labeling pipeline

To build a model capable of verifying both object type and condition, we developed a tailored dataset from high-resolution images captured directly by the client’s machines. Since existing public datasets lacked the precision or licensing required, we established a custom ingestion pipeline to organize raw image archives, eliminate duplicates, and preserve metadata such as object dimensions and material.

We deployed a self-hosted instance of Label Studio and onboarded an external labeling team to efficiently annotate thousands of samples. Clear labeling guidelines and active feedback loops ensured accurate segmentation, even in edge cases like open bottle caps or pull-tabs. This labeling infrastructure was designed for reusability—allowing the client to expand the dataset and improve accuracy over time.

Iterative model development for segmentation and verification

We evaluated existing image segmentation models but found them too large or slow for on-device execution. Instead, our team took an iterative approach: a data scientist built an initial segmentation model optimized for quality, which was then fine-tuned, ported, and tested by an ML engineer for performance on the target hardware.

Each iteration was tracked using Neptune.ai, with benchmark metrics including per-image inference time and hardware compatibility via ONNX Runtime. This approach allowed us to gradually reduce model complexity while preserving accuracy, ultimately delivering a version that met both speed and reliability thresholds.

Lightweight, deployable architecture

To simplify installation and updates, we shipped the entire solution as a versioned Python package (.tar.gz), including:

This modular packaging strategy reduced update sizes and made it easy for the client to deploy the solution at scale across multiple machines, even in bandwidth-limited environments.

Ready for future expansion

The system architecture supports future needs like:

The final solution delivers accurate, real-time image analysis at the edge, enabling secure, fraud-resistant recycling flows, with no cloud connectivity required.

Case Study Schema Building a Real-Time Packaging Recognition System with Edge-Optimized AI

Tools

Python

Python

Google Cloud Platform

Google Cloud Platform

Have a similar project in mind?

LET'S TALK!

Challenges

Fragmented and low-quality image data
The first batches of imagery provided by the client were heavily downsampled, making it impossible to train a reliable model for shape or material detection. Without access to raw image detail, both segmentation accuracy and object classification suffered. We had to work with the client to re-establish a data acquisition process that delivered high-resolution images consistently and with proper metadata.

Lack of suitable public datasets
Off-the-shelf datasets fell short, not due to class coverage or volume alone, but because they lacked the specific object positioning and framing required for this use case. Since the system had to verify not just the type of container but also its shape and dimensions, precision was critical. Public datasets like TACO or the Bottles Dataset didn’t offer consistently positioned, high-resolution images that would support reliable size estimation or edge-case detection. To meet these specific requirements, we designed and built a custom dataset pipeline, from high-resolution image ingestion and deduplication to segmentation labeling and metadata tracking, ensuring full control over data quality and consistency.

Edge deployment under strict latency constraints
Running deep learning models on a Raspberry Pi 5—within a 500ms budget—required precise engineering. Most existing segmentation models were too large or compute-intensive. We couldn't rely on batching or cloud offloading, so the solution had to be lean and local. Our team iterated between quality tuning and ONNX-based optimization to reach the performance envelope, making careful trade-offs between speed and accuracy.

Ambiguity in packaging edge cases
Real-world recyclables aren't always perfect. Some items came with unscrewed bottle caps, crushed necks, or partially torn labels, leading to labeling inconsistencies. Rather than over-engineering logic for every corner case, we created clear annotation standards and maintained a feedback loop with the external labeling team. This helped enforce consistency across the dataset and reduced noise during training.

Scalability and updatability across devices
With hundreds (and eventually thousands) of machines expected to run the model, maintenance needed to be lightweight. Static deployment packages quickly became impractical. We designed a packaging approach that kept the Python module and requirements separate, allowing the client to push targeted updates without re-downloading the entire environment.

Our other projects:

Fintech Sector

Loan Application Fraud Detection

Developing a model that can identify fraudulent loan applications for a financial institution?
Challenge Accepted!

Fraud Detection Predictive Modeling

Orkla Home & Personal Care (OHPC)

Optimizing Procurement through Seamless RFQ Management

The new AI-powered procurement system reduced RFQ processing time by up to 85%, cutting the cycle from 4 weeks to just 4 days and enabling faster, more cost-efficient purchasing decisions.

Challenge Accepted!

Data Engineering Generative AI LLMs

Insurance Sector

Auto Insurance Fraud Claim Detection

Developing an AI model that can detect car insurance fraud based on car accident claims?
Challenge Accepted!

Fraud Detection Predictive Modeling

Fintech Sector

Loan Application Fraud Detection

Developing a model that can identify fraudulent loan applications for a financial institution?
Challenge Accepted!

Fraud Detection Predictive Modeling

Orkla Home & Personal Care (OHPC)

Optimizing Procurement through Seamless RFQ Management

The new AI-powered procurement system reduced RFQ processing time by up to 85%, cutting the cycle from 4 weeks to just 4 days and enabling faster, more cost-efficient purchasing decisions.

Challenge Accepted!

Data Engineering Generative AI LLMs

See more projects

This site uses cookies for analytical purposes.

Accept Privacy Policy

In the interests of your safety and to implement the principle of lawful, reliable and transparent processing of your personal data when using our services, we developed this document called the Privacy Policy. This document regulates the processing and protection of Users’ personal data in connection with their use of the Website and has been prepared by Nexocode.

To ensure the protection of Users' personal data, Nexocode applies appropriate organizational and technical solutions to prevent privacy breaches. Nexocode implements measures to ensure security at the level which ensures compliance with applicable Polish and European laws such as:

  1. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation) (published in the Official Journal of the European Union L 119, p 1); Act of 10 May 2018 on personal data protection (published in the Journal of Laws of 2018, item 1000);
  2. Act of 18 July 2002 on providing services by electronic means;
  3. Telecommunications Law of 16 July 2004.

The Website is secured by the SSL protocol, which provides secure data transmission on the Internet.

1. Definitions

  1. User – a person that uses the Website, i.e. a natural person with full legal capacity, a legal person, or an organizational unit which is not a legal person to which specific provisions grant legal capacity.
  2. Nexocode – NEXOCODE sp. z o.o. with its registered office in Kraków, ul. Wadowicka 7, 30-347 Kraków, entered into the Register of Entrepreneurs of the National Court Register kept by the District Court for Kraków-Śródmieście in Kraków, 11th Commercial Department of the National Court Register, under the KRS number: 0000686992, NIP: 6762533324.
  3. Website – website run by Nexocode, at the URL: nexocode.com whose content is available to authorized persons.
  4. Cookies – small files saved by the server on the User's computer, which the server can read when when the website is accessed from the computer.
  5. SSL protocol – a special standard for transmitting data on the Internet which unlike ordinary methods of data transmission encrypts data transmission.
  6. System log – the information that the User's computer transmits to the server which may contain various data (e.g. the user’s IP number), allowing to determine the approximate location where the connection came from.
  7. IP address – individual number which is usually assigned to every computer connected to the Internet. The IP number can be permanently associated with the computer (static) or assigned to a given connection (dynamic).
  8. GDPR – Regulation 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of individuals regarding the processing of personal data and onthe free transmission of such data, repealing Directive 95/46 / EC (General Data Protection Regulation).
  9. Personal data – information about an identified or identifiable natural person ("data subject"). An identifiable natural person is a person who can be directly or indirectly identified, in particular on the basis of identifiers such as name, identification number, location data, online identifiers or one or more specific factors determining the physical, physiological, genetic, mental, economic, cultural or social identity of a natural person.
  10. Processing – any operations performed on personal data, such as collecting, recording, storing, developing, modifying, sharing, and deleting, especially when performed in IT systems.

2. Cookies

The Website is secured by the SSL protocol, which provides secure data transmission on the Internet. The Website, in accordance with art. 173 of the Telecommunications Act of 16 July 2004 of the Republic of Poland, uses Cookies, i.e. data, in particular text files, stored on the User's end device.
Cookies are used to:

  1. improve user experience and facilitate navigation on the site;
  2. help to identify returning Users who access the website using the device on which Cookies were saved;
  3. creating statistics which help to understand how the Users use websites, which allows to improve their structure and content;
  4. adjusting the content of the Website pages to specific User’s preferences and optimizing the websites website experience to the each User's individual needs.

Cookies usually contain the name of the website from which they originate, their storage time on the end device and a unique number. On our Website, we use the following types of Cookies:

  • "Session" – cookie files stored on the User's end device until the Uses logs out, leaves the website or turns off the web browser;
  • "Persistent" – cookie files stored on the User's end device for the time specified in the Cookie file parameters or until they are deleted by the User;
  • "Performance" – cookies used specifically for gathering data on how visitors use a website to measure the performance of a website;
  • "Strictly necessary" – essential for browsing the website and using its features, such as accessing secure areas of the site;
  • "Functional" – cookies enabling remembering the settings selected by the User and personalizing the User interface;
  • "First-party" – cookies stored by the Website;
  • "Third-party" – cookies derived from a website other than the Website;
  • "Facebook cookies" – You should read Facebook cookies policy: www.facebook.com
  • "Other Google cookies" – Refer to Google cookie policy: google.com

3. How System Logs work on the Website

User's activity on the Website, including the User’s Personal Data, is recorded in System Logs. The information collected in the Logs is processed primarily for purposes related to the provision of services, i.e. for the purposes of:

  • analytics – to improve the quality of services provided by us as part of the Website and adapt its functionalities to the needs of the Users. The legal basis for processing in this case is the legitimate interest of Nexocode consisting in analyzing Users' activities and their preferences;
  • fraud detection, identification and countering threats to stability and correct operation of the Website.

4. Cookie mechanism on the Website

Our site uses basic cookies that facilitate the use of its resources. Cookies contain useful information and are stored on the User's computer – our server can read them when connecting to this computer again. Most web browsers allow cookies to be stored on the User's end device by default. Each User can change their Cookie settings in the web browser settings menu: Google ChromeOpen the menu (click the three-dot icon in the upper right corner), Settings > Advanced. In the "Privacy and security" section, click the Content Settings button. In the "Cookies and site date" section you can change the following Cookie settings:

  • Deleting cookies,
  • Blocking cookies by default,
  • Default permission for cookies,
  • Saving Cookies and website data by default and clearing them when the browser is closed,
  • Specifying exceptions for Cookies for specific websites or domains

Internet Explorer 6.0 and 7.0
From the browser menu (upper right corner): Tools > Internet Options > Privacy, click the Sites button. Use the slider to set the desired level, confirm the change with the OK button.

Mozilla Firefox
browser menu: Tools > Options > Privacy and security. Activate the “Custom” field. From there, you can check a relevant field to decide whether or not to accept cookies.

Opera
Open the browser’s settings menu: Go to the Advanced section > Site Settings > Cookies and site data. From there, adjust the setting: Allow sites to save and read cookie data

Safari
In the Safari drop-down menu, select Preferences and click the Security icon.From there, select the desired security level in the "Accept cookies" area.

Disabling Cookies in your browser does not deprive you of access to the resources of the Website. Web browsers, by default, allow storing Cookies on the User's end device. Website Users can freely adjust cookie settings. The web browser allows you to delete cookies. It is also possible to automatically block cookies. Detailed information on this subject is provided in the help or documentation of the specific web browser used by the User. The User can decide not to receive Cookies by changing browser settings. However, disabling Cookies necessary for authentication, security or remembering User preferences may impact user experience, or even make the Website unusable.

5. Additional information

External links may be placed on the Website enabling Users to directly reach other website. Also, while using the Website, cookies may also be placed on the User’s device from other entities, in particular from third parties such as Google, in order to enable the use the functionalities of the Website integrated with these third parties. Each of such providers sets out the rules for the use of cookies in their privacy policy, so for security reasons we recommend that you read the privacy policy document before using these pages. We reserve the right to change this privacy policy at any time by publishing an updated version on our Website. After making the change, the privacy policy will be published on the page with a new date. For more information on the conditions of providing services, in particular the rules of using the Website, contracting, as well as the conditions of accessing content and using the Website, please refer to the the Website’s Terms and Conditions.

Nexocode Team