Rapporten presenterar det arbete som utfördes för en kandidatuppsats i ämnesområdet datavetenskap. Den ursprungliga uppgiften var att undersöka hur djupinlärningsarkitekturen ESRGAN, som används för superupplösning, kan komprimeras så att minimal precision förloras. Projektet resulterade i utvärderingen av tre optimeringsmetoder; dynamic range, full integer och float16-kvantisering. Mätningarna utfördes med hjälp av två mobila enheter; en Samsung Galaxy S9+ surfplatta och en S10+ Android-telefon. Mätningarna genomfördes med hjälp av mätvärdena inferenstid, PSNR, SSIM och kompressionsförhållande. Resultaten visade att Dynamic Range hade en avsevärt långsammare inferenstid jämfört med Full Integer och Float16-kvantisering. Dynamic Range hade ett validerings-PSNR på 27.0 och ett test-PSNR på 22.3. De resulterande SSIM-värdena var 0.81 för valideringsdatasetet och 0.67 för testdatasetet. Full Integer slutade med PSNR-värdena 26.3 och 21.9 för validering respektive test. När det gäller SSIM fick Full Integer poängen 0.77 (validering) och 0.64 (test). Slutligen genererade Float16 PSNR-värdena 27.1 och 22.3, samt SSIM-värdena 0.81 och 0.67. PSNR- och SSIM-utvärderingarna visade att de komprimerade modellerna behövde mer kalibrering för att uppnå högre poäng i dessa metoder, och således högre noggrannhet. / This report presents the work that was carried out for a bachelor’s thesis in computer science. The original task was to investigate how the deep learning architecture ESRGAN used for super resolution can be compressed such that minimal accuracy is lost. The project resulted in the evaluation of three optimization methods; dynamic range, full integer, and float16 quantization. Dynamic range quantizes the weights of the neural network into 8 bits of precision, full integer quantizes all floating point parameters, and float16 reduces halves the floating point precisions. The benchmarks were performed using two mobile devices; a Samsung Galaxy S9+ tablet and an S10+ android phone. Measurements were conducted using metrics inference time, PSNR, SSIM, and compression ratio. The results showed that Dynamic Range had a significantly slower inference time compared to Full Integer and Float16 quantization. Dynamic range had the validation PSNR score of 27.0, and a testing PSNR score of 22.3. The resulting SSIM values were 0.81 for the validation dataset and 0.67 for the testing dataset. Full integer ended up with the PSNR scores 26.3, 21.9 for validation and testing respectively. As for SSIM, Full integer brought the scores 0.77 (validation) and 0.64 (testing). Finally, Float16 generated PSNR scores 27.1 and 22.3, and the SSIM scores 0.81 and 0.67. The PSNR and SSIM evaluations showed that the compressed models needed more calibration for a higher score in these metrics, and consequently a higher level of accuracy.

Game engines such as Unreal Engine 5 are widely used to create photo-realistic renderings. To run these renderings at high quality without experiencing any performance issues,high-performance hardware is often required. In situations where the hardware is lacking,users may be forced to lower the quality and resolution of renderings to maintain goodperformance. While this may be acceptable in some situations, it limits the benefit that apowerful tool like Unreal Engine 5 can provide. This thesis aims to explore the possibilityof using a Real-ESRGAN, fine-tuned on a custom data set, to increase both the resolutionand quality of screenshots taken in Unreal Engine 5. By doing this, users can lower theresolution and quality of their Unreal Engine 5 rendering while still being able to generatehigh quality screenshots similar to those produced when running the rendering at higherresolution and higher quality. To accomplish this, a custom data set was created by randomizing camera positionsand capturing screenshots in an Unreal Engine 5 rendering. This data set was used to finetune a pre-trained Real-ESRGAN model. The fine-tuned model could then generate imagesfrom low resolution and low quality screenshots taken in Unreal Engine 5. The resultingimages were analyzed and evaluated using both quantitative and qualitative methods.The conclusions drawn from this thesis indicate that images generated using the finetuned weights are of high quality. This conclusion is supported by quantitative measurements, demonstrating that the generated images and the ground truth images are similar.Furthermore, visual inspection conducted by the authors confirms that the generated images are similar to the reference images, despite occasional artifacts.

This thesis investigates the impact of image enhancement techniques on object detection for carsin real-world traffic scenarios. The study focuses on upscaling and light correction treatments andtheir effects on detecting cars in challenging conditions. Initially, a YOLOv8x model is trained on clear static car images. The model is then evaluated on a test dataset captured in real-world driving with images from a front-mounted camera on a car, incorporating various lighting conditions and challenges. The images are then enhanced with said treatments and then evaluated again. The results in this experiment with its specific context show that upscaling seems to decreasemAP performance while lighting correction slightly improves accuracy. Additional training on acomplex image dataset outperforms all other approaches, highlighting the importance of diverse and realistic training data. These findings contribute to advancing computer vision research for object detection models.