IMAGE SIMILARITY MEASUREMENT USING RESNET50 AND STREET VIEW ANALYSIS FOR RESIDENTIAL LOCATION POINT IDENTIFICATION

Nadhief, Naufal (2026) IMAGE SIMILARITY MEASUREMENT USING RESNET50 AND STREET VIEW ANALYSIS FOR RESIDENTIAL LOCATION POINT IDENTIFICATION. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 19081010056_cover.pdf Download (769kB)
	Text (Bab 1) 19081010056_bab1.pdf Download (223kB)
	Text (Bab 2) 19081010056_bab2.pdf Restricted to Repository staff only until 21 July 2029. Download (569kB)
	Text (Bab 3) 19081010056_bab3.pdf Restricted to Repository staff only until 21 July 2029. Download (533kB)
	Text (Bab 4) 19081010056_bab4.pdf Restricted to Repository staff only until 21 July 2029. Download (2MB)
	Text (Bab 5) 19081010056_bab5.pdf Download (186kB)
	Text (Daftar Pustaka) 19081010056_daftarpustaka.pdf Download (188kB)

Abstract

Building asset management in Bojonegoro Regency still faces challenges due to the lack of visual documentation integrated with spatial data. Manual field surveys require significant costs, time, and labor, while the available Google Street View imagery exhibits a high level of redundancy due to image capture at closely spaced locations. This study proposes an automated deep learning-based system to identify building locations using Google Street View imagery. The method utilizes ResNet50 as a feature extractor to generate 2048-dimensional feature vectors from each image, followed by the application of cosine similarity to measure the degree of resemblance between images so that redundant images can be eliminated. The research data were collected from the Kapas, Balen, and Dander subdistricts in Bojonegoro Regency. Testing across four threshold variations (0.85, 0.90, 0.95, and 0.98) demonstrated that a threshold of 0.95 provided the best balance by detecting 112 redundant images with only 2 elimination errors. Per-subdistrict similarity testing yielded cosine similarity values ranging from 0.83 to 0.94, indicating that all test image pairs fell below the threshold, meaning no images were incorrectly eliminated. Furthermore, classification testing on 30 sample images of residential houses and shops produced confidence values between 97.6% and 99.4%, demonstrating the model's ability to recognize visual characteristics of buildings with a high degree of confidence. The geographic coordinates of the remaining images were subsequently integrated with the Google Maps API through a reverse geocoding process to automatically obtain building location addresses. The results indicate that the proposed method is capable of producing more efficient representative images and supports automated building address identification for regional asset management purposes

Item Type:

Thesis (Undergraduate)

Contributors: