A Modular Approach for Object Segmentation and Image Inpainting

Silva, Alexandre Santos

http://hdl.handle.net/10400.8/14628

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
ProjectReport_AlexandreSantosSilva_2230329_c_f.pdf		66.67 MB	Adobe PDF	Download

Send Feedback

Authors

Silva, Alexandre Santos

Advisor(s)

Ribeiro, José Carlos Bregieiro

Miragaia, Rolando Lúcio Germano

Silva, Fernando José Mateus da

Costa, Rogério Luís de Carvalho

Abstract(s)

In this project, we propose ModInPainTor, a modular solution for object segmentation and removal in images. ModInPainTor takes advantage of advanced models from literature to deliver high-quality image inpainting results. This solution is inspired by InPainTor, a previous project that utilized a highly-integrated Convolutional Neural Network-based model for object detection and removal but lacked a comprehensive evaluation of its performance. We conduct a thorough evaluation of the InPainTor and ModInPainTor solutions, assessing their effectiveness in anonymizing visual data while maintaining data quality. The evaluation involves quantitative metrics to measure the quality of the results and a qualitative analysis to evaluate the visual quality of the anonymized images. Furthermore, we investigate the impact of object segmentation accuracy on the quality of the anonymized images. Our findings indicate that Mod- InPainTor significantly improves the visual quality of anonymized images compared to InPainTor, albeit with increased computational requirements. We find that the proposed solution could be effectively employed in industrial settings where anonymization quality is prioritized over computational efficiency.