Electronic Theses and Dissertation

Penelitian ini bertujuan untuk merancang arsitektur deep learning dengan komputasi rendah yang mampu melakukan pencarian pakaian menggunakan citra sketsa, serta menguji dan membandingkan kemampuan model-model deep learning seperti shufflenet v2, mobilenet v2, dan efficientnet-b1 dalam temu kembali citra pakaian berbasis sketsa. metode yang diimplementasikan menggabungkan ekstraksi fitur menggunakan discrete wavelet transform (dwt) dengan model-model deep learning tersebut. pengujian dilakukan dengan pencarian terhadap dataset foto pakaian menggunakan citra sketsa pakaian sebagai masukan terhadap 17 kategori pakaian yaitu bermuda, blouse, cardigan, circle, dress, hoodie, jacket, jumpsuit, mini, pencil, shirt, skinny, straight, suit, t-shirt, wideleg, dan tunic. hasil penelitian menunjukkan bahwa penggunaan dwt meningkatkan kinerja model dibandingkan dengan model tanpa dwt. model efficientnet-b1 yang menggunakan dwt mengungguli mobilenet v2 dan shufflenet v2 yang juga menggunakan dwt. peningkatan kinerja diukur berdasarkan nilai precision@5, recall@5, dan mean average precision (map). untuk shufflenet v2, penggunaan dwt meningkatkan nilai precision@5, recall@5, dan map masing-masing sebesar 5.71%, 10.77%, dan 7.35%. untuk mobilenet v2, peningkatan masing-masing sebesar 4.21%, 4.29%, dan 4.23%. efficientnet-b1 menunjukkan peningkatan masing-masing sebesar 6.85%, 7.04%, dan 6.94%

The purpose of this research is to design a low-computation deep learning architecture capable of performing searches for clothes using sketch images, as well as to test and compare the capabilities of deep learning models such as ShuffleNet v2, MobileNet v2, and EfficientNet-B1 in sketch-based dress image retrieval. The method implemented combines feature extraction using the Discrete Wavelet Transform (DWT) with deep learning models. We tested the models by searching for clothing photos using sketches as input for 17 clothing categories: bermuda, blouse, cardigan, circle, dress, hoodie, jacket, jumpsuit, mini, pencil, shirt, skinny, straight, suit, t-shirt, wide-leg, and tunic. The results showed that using DWT improved model performance. The EfficientNet-B1 model, which uses DWT, outperformed MobileNet v2 and ShuffleNet v2, which also use DWT. We measured performance improvements based on Precision@5, Recall@5, and mean Average Precision (mAP). For ShuffleNet v2, DWT increased the values of Precision@5, Recall@5, and mAP by 5.71%, 10.77%, and 7.35%. For MobileNet v2, the improvements were 4.21%, 4.29%, and 4.23%, and for EfficientNet-B1 the improvements were 6.85%, 7.04%, and 6.94%.