Datamatrix (DM) codes – similar to QR codes – are compact visual text-encodings that are heavily used in industry to label industrial products or packages. The objective of this work is to adopt neural networks (NN) to enhance the decodability of low-visible or otherwise visually damaged images of DM codes. In the theoretical part, the student will study the domain of DM codes described in the GS1 DM standard. Suitable deep NN encoder-decoder architectures (e.g., Unet) for image enhancement will be described. In the practical part, the student will train networks that take an image containing a DM code in bad quality (distorted, blurred, etc.) on the input and provide as-readable-as-possible filtered output. The influence of the selection of encoders, network architectures, and training parameters will be examined and measured. Accuracy (decode rate, misread rate) and performance (inference time) of selected approaches will be reported. The student will create training, validation, and test datasets. Training data will be synthetically generated using an open-source DM generator combined with suitable image augmentations. Test data will be acquired from real-world photos.