Electronic Theses and Dissertation

Ikan merupakan bahan pangan yang mudah mengalami penurunan kesegaran (perishable food), sehingga penentuan tingkat kesegarannya menjadi hal penting dalam menjaga kualitas dan keamanan pangan. penilaian kesegaran ikan secara manual melalui pengamatan visual terhadap mata ikan masih bersifat subjektif dan bergantung pada pengalaman pengamat, sehingga diperlukan pendekatan otomatis berbasis citra digital. penelitian ini bertujuan membandingkan performa model mobilenetv2 dengan pendekatan transfer learning dan metode stacking ensemble learning yang memanfaatkan mobilenetv2 sebagai ekstraktor fitur dalam mengklasifikasikan kesegaran ikan berdasarkan citra mata ikan, menggunakan dataset fish freshness dari roboflow universe yang terbagi atas kelas fresh dan not fresh dengan proporsi 80% data latih, 10% data validasi, dan 10% data uji. evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan f1-score. hasil pengujian menunjukkan bahwa mobilenetv2 memperoleh performa yang lebih baik dengan nilai accuracy sebesar 87,31%, precision 0,87, recall 0,87, dan f1-score 0,87, dibandingkan stacking ensemble learning yang memperoleh accuracy 81,14%, precision 0,85, recall 0,81, dan f1-score 0,80, serta jumlah false negative yang lebih rendah. model mobilenetv2 terbaik selanjutnya diimplementasikan ke dalam prototipe aplikasi web berbasis gradio yang memungkinkan pengguna mengunggah citra mata ikan dan memperoleh hasil klasifikasi kesegaran ikan secara otomatis beserta nilai confidence score-nya. kata kunci: mobilenetv2, stacking ensemble learning, klasifikasi kesegaran ikan, citra mata ikan, transfer learning

Fish is a perishable food commodity that can easily degrade in freshness, making the determination of its freshness level an important aspect of maintaining food quality and safety. Manual freshness assessment through visual observation of fish eyes remains subjective and dependent on the observer's experience, highlighting the need for an automated approach based on digital imagery. This study aims to compare the performance of the MobileNetV2 model using a transfer learning approach with the Stacking Ensemble Learning method, which utilizes MobileNetV2 as a feature extractor, in classifying fish freshness based on fish eye images, using the Fish Freshness dataset from Roboflow Universe divided into fresh and not fresh classes with a split of 80% training data, 10% validation data, and 10% testing data. Evaluation was carried out using accuracy, precision, recall, and F1-score metrics. The test results show that MobileNetV2 achieved better performance, with an accuracy of 87.31%, precision of 0.87, recall of 0.87, and F1-score of 0.87, compared to Stacking Ensemble Learning, which achieved an accuracy of 81.14%, precision of 0.85, recall of 0.81, and F1-score of 0.80, along with a lower number of false negatives. The best-performing MobileNetV2 model was then implemented into a Gradio-based web application prototype that allows users to upload fish eye images and automatically obtain fish freshness classification results along with their confidence scores. Keywords: MobileNetV2, Stacking Ensemble Learning, fish freshness classification, fish eye image, transfer learning