Image Geolocalization via Combined Classification and Regression

This paper presents a geolocation system to assist in identifying the location in images associated with illicit activities, thereby reducing investigation time for authorities. The system is implemented as a multi-agent system. The geolocalisation is performed based on image analysis, which involves both image classification and regression of the geographical coordinates. The final result is obtained by fusioning both results. The classification model is based on the EfficientNet architecture. The evaluation is performed on a dataset with images from Romania collected from Google Street View. Moreover, the best version of the model succeeds in solving difficult classification tasks, correctly identifying the answer among two or more similar counties in 80% of the cases. Alternatively, from a regression point of view, the predicted coordinates are at most 80 km away from the ground truth location in over 80% of the evaluation cases. By employing various data augmentation techniques, such as noise addition or severe compression, a model robust to potential adversarial interference was achieved.