This thesis aims to update the Android application TUM-Lens by loading the deep learning models dynamically and doing the image analysis (image classification and object detection) locally on the smartphone. It also uses the concept of transfer learning in order to feed the model zoo with many more models. The dynamic loading concept leads to an empirical minimization of the app size. In addition, we add other features to the app, such as adding a delay in which the user can change. For feeding the model zoo, we use transfer learning. In this work, we use five different models that are pre-trained on the ImageNet dataset; the models used are (MobileNet, Inception, ResNet, NasNet, and Inception ResNet). We also train these models with three optimizers (Adam, SGD, and Newton). These models were used against two datasets (Intel dataset and Cinic-10 dataset). For each dataset, we use all the possible combinations between models and optimizers and monitor the accuracy of each case. In the end, we came up with three different models with high accuracy and added them to the model zoo of the app to be used in the image classification mode. Moreover, we add one TensorFlow lite model to the object detection mode.     «

This thesis aims to update the Android application TUM-Lens by loading the deep learning models dynamically and doing the image analysis (image classification and object detection) locally on the smartphone. It also uses the concept of transfer learning in order to feed the model zoo with many more models. The dynamic loading concept leads to an empirical minimization of the app size. In addition, we add other features to the app, such as adding a delay in which the user can change. For fee...     »