Master thesis : Research of an appropriate neural network architecture for the fast and accurate detection of DataMatrixCode

Abstract

In a world where more and more progress is made in various fields using the computational power of modern computers, many industries are using barcodes for different purposes going from identifying products to retrieve information to locating oneself in areas of a warehouse. Decoding barcodes is thus at the core of processes of certain companies. Moreover, this task has to be done effectively, in terms of time and accuracy, making it challenging in real world applications where the barcode may not be at the center of the image, partly occluded, distorted or very blurry. Classical computer vision methods and models can be effective for certain types of images but when conditions are poor, they don't seem to work that fine. Fortunately, deep learning especially in the computer vision research field has been having great success among the researchers, making the progress in object detection task better and better. This work aims at finding a neural network architecture capable of locating Data Matrix code at a high speed with great accuracy. This work is mainly based around You Only Look Once (Yolo) and its upgrades throughout the years, while exploring different networks for optimal speed. Firstly, Data Matrix code will be detected using bounding boxes with their axis parallel to the borders of the images and later on, we will detect them with their four corners to make it easier for a decoding gent to identify the code, while having small and rapid models able to be run on CPU.