We have developed a solution to identify the information printed on packages, execute on small hardware, with high processing speed and accuracy.
In warehouse management, incoming package should be pre-checked by the workers. There are two types of information to be checked. The first type is fixed information such as product name, place of manufacture, ingredient. This information is usually pre-printed and encoded in the product barcode. The second type is non-fixed information such as the use-by date. This information will be printed on the box after the product is packaged.
Our solution is capable of reading information on products, including fixed (via barcode) and non-fixed. The solution uses a system consisting of multiple cameras mounted around the conveyor belt. Imcoming parcels are placed on conveyor belts to check information. The recognition engine can be executed on small hardware like the Jetson TX2 and still achieves the processing speed of 2 seconds per product.
Our solution is a combination of fine-tuned text detection and recognition models, regex-based content analysis. Text-processing models are fine-tuned to handle special font formats for printing on packaging such as dotted fonts