Finding a way to recognize people from their sketches is of great importance for law enforcement. Automatic retrieval of photos of suspects from police mug-shot databases using sketches drawn by artists can quickly help police to narrow down potential suspects and facilitate the investigation process. In this thesis, we have developed a deep learning model based on a famous Siamese Network architecture, which takes a sketch and a photo as input and gives out a similarity score as an output using deep features. 

The initial dataset comprises two images: one depicting a sketch and the other representing the corresponding actual photo. Due to the limited availability of sketch images, we employed cycle GAN (refer to Figure 4.3) to generate additional sketches for training purposes, as illustrated in Figure 4.6. Finally, the last figure showcases the test results, displaying the dissimilarity score between the base image and the other images.

In the thesis project, I played a pivotal role as part of a team of three students. My responsibilities encompassed collecting and preprocessing the data, along with contributing to the design of the model. Additionally, I am pleased to mention that I was awarded a scholarship for my work and academic performance. 

Counterfeit banknotes have significant negative effects on a country's economy and society, leading to reduced value of real money, increased inflation, and potential recessions. In this study, we proposed a Perceptual Autoencoder model to detect counterfeit banknotes as an anomaly detection problem, training the model on one class. We also implemented a RestNet18-based model as a baseline, fine-tuned for the task. Using publicly available datasets and a custom Ethiopian banknote dataset, we evaluated the models using the AUC-ROC metric. The ResNet18 model outperformed the perceptual Autoencoder model, achieving average AUC-ROC scores of 99.99% and 98.63%, respectively. The Autoencoder model had longer training times, suggesting the need for optimization. Future work should focus on reducing training time for the Autoencoder model and incorporating explainable AI techniques for both models to enhance prediction interpretability.

The first two images are the data preprocessing steps and the sample dataset samples. The image in Figure 3.14 is the proposed model for the counterfeit detection problem. The last image set is the reconstructed images for the first 1600 iterations of training.

This research project, conducted at my previous university, was guided by a senior researcher and funded by the institution. It took place during the year when the government introduced new banknotes, addressing a significant issue of counterfeiting. Our aim was to find a solution despite the limited availability of training negative samples (counterfeited banknotes).