Reidentifikasi Orang pada Data Parsial menggunakan Klasifier Swin Transformer

Soemarso, Bayuadi Prakoso (2024) Reidentifikasi Orang pada Data Parsial menggunakan Klasifier Swin Transformer. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024201075-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (12MB) | Request a copy

Abstract

Penelitian ini dilatarbelakangi oleh kebutuhan untuk mengembangkan sistem reidentifikasi orang yang efektif dalam situasi di mana hanya data parsial tersedia, seperti dalam pengawasan keamanan di area publik atau lalu lintas. Tantangan ini mencakup identifikasi individu dalam kondisi pencahayaan rendah atau cakupan kamera terbatas. Metode yang digunakan dalam penelitian ini melibatkan persiapan dataset Partial-REID dan Market1501, pengembangan model Swin Transformer, pelatihan data dengan hyperparameter yang telah ditentukan, pengujian data, dan evaluasi model. Hasil penelitian menunjukkan bahwa model Swin Transformer yang dikembangkan mencapai hasil training loss sekitar 0.02 dan akurasi sekitar 0.99, serta validasi loss sekitar 0.005 dan akurasi 0.98. Pengujian model menghasilkan Rank@1 sebesar 0.79, Rank@5 sebesar 0.89, Rank@10 sebesar 0.93, dan mAP sebesar 0.73. Setelah re-ranking menggunakanK-ReciprocalRe-Rank, hasilnyameningkatmenjadiRank@1sebesar0.75, Rank@5 sebesar 0.81, Rank@10 sebesar 0.88, dan mAP sebesar 0.76. Kesimpulannya, model Swin Transformer berhasil meningkatkan kinerja reidentifikasi orang pada data parsial, menunjukkan akurasi yang tinggi dan kemampuan adaptasi yang baik terhadap variasi kondisi pengambilan gambar.
===================================================================================================================================
This research is motivated by the need to develop an effective person re-identification system in situations where only partial data is available, such as in security surveillance in public areas or traffic. These challenges include identifying individuals in low lighting conditions or limited camera coverage. The method used in this research involves preparing the Partial-REID and Market1501 datasets, developing the Swin Transformer model, training data with predetermined hyperparameters, testing data, and evaluating the model. The research results show that the Swin Transformer model developed achieved training loss of around 0.02 and accuracy of around 0.99, as well as validation loss of around 0.005 and accuracy of 0.98. Model testing produces Rank@1 of 0.79, Rank@5 of 0.89, Rank@10 of 0.93, and mAP of 0.73. After re-ranking using K-Reciprocal Re-Rank, the results increased to Rank@1 of 0.75, Rank@5 of 0.81, Rank@10 of 0.88, and mAP of 0.76. In conclusion, the Swin Transformer model successfully improves person re-identification performance on partial data, showing high accuracy and good adaptability to variations in shooting conditions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Swin Transformer, Reidentifikasi-orang, Data Parsial, Deep Learning, Pengenalan gambar; Swin Transformer, Person re-identification, Partial data, Deep learning, Image Recognition.
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Bayuadi Prakoso Soemarso
Date Deposited:	26 Jul 2024 01:09
Last Modified:	26 Jul 2024 01:09
URI:	http://repository.its.ac.id/id/eprint/108949

Actions (login required)

View Item