About the Company:

The customer is a a cutting-edge 360 degree “Design-to-Deliver” company redefining the supply innovation of automotive & industrial parts & components.

They cover the entire spectrum of the supply chain right from cloud manufacturing, dynamic fulfilment, quality control and last mile delivery. Their tech platform empowers them to assess the performance capabilities of vast network of suppliers and forge a dynamic ecosystem of cloud-based manufacturers along with providing visibility and traceability throughout the value chain.

Problem Statement:

The customer sources printed stickers and user manuals from their vendors, and supplies them to their customers to be used on electronic appliances. Previously, the quality control of these stickers was done by a human, validating if everything was printed properly. However, the chance of a human making an error while validating text is high when working at high speeds.

Every error for the customer was a potential lawsuit by their customers, since that would mean misinforming the customers on the appliance specifications. Hence, it was important to catch 100% of printing errors through a systematic solution.

Key Challenges:

Out of the box solutions like using an OCR does not work, as it has only 95% accuracy in reading text. Moreover, the printed text often includes finer fonts and logos that are not caught by OCR. Additionally, it images are captured by a human, and contain deviation in camera alignment, lighting, product packaging, etc. Hence, a threshold must be considered to eliminate noise, yet, detect all the errors.

Approach:

To fit this into the existing production process of the customer, we decided to build a mobile application that could capture images of products, and highlight the defects within the UI. Additionally, it would also send a consolidated report of the findings to be visualized on a dashboard.

To accurately identify 100% of the defects, we used two ML models, the first being pixel matching and the second being OCR. Either one of the models did not provide satisfactory results when used in isolation, and hence, we decided to combine the two to achieve the outcome.

Our solution was able to detect presence of text and the written content with OCR, and scanned the same with our pixel matching algorithm as well. If the combined metric was satisfactory, then it wasn’t a defect, else it was marked as a defect, and shown in the form of a bounding box in the UI.

Using Pixel matching required considerable pre-processing in the form of resizing, warping, color correcting, etc., without which there was no chance of using the method.

‍