Electronic Theses and Dissertation

Deteksi ikan bawah air penting untuk pemantauan ekosistem laut, namun kinerjanya sering menurun akibat degradasi citra seperti kontras rendah, dominasi spektrum biru–hijau, kekeruhan, dan variasi pencahayaan. penelitian ini mengusulkan pipeline deteksi ikan bawah air yang mengintegrasikan peningkatan citra berbasis clahe dan deteksi objek menggunakan yolov3, serta memvalidasi implementasinya pada perangkat edge ai nvidia jetson orin nano melalui antarmuka streamlit. empat varian peningkatan citra dievaluasi, yaitu clahe, clahe+unsharp masking (usm), clahe+high-frequency emphasis filter (hef), dan clahe+percentile blending. evaluasi kualitas citra dilakukan pada subset 10% dataset deepfish menggunakan metrik loe, uiqm, dan uciqe. hasil menunjukkan clahe+percentile blending memberikan kompromi terbaik dengan loe terendah (212,021) serta uiqm (3,949) dan uciqe (0,508) tertinggi. evaluasi deteksi dilakukan pada dua domain: dataset deepfish dan lokal sabang, aceh. pada deepfish, peningkatan citra meningkatkan map@0,5 dari 96,15% menjadi 97,05%, menurunkan false negative (198 menjadi 137), dan meningkatkan iou rata-rata (74,39% menjadi 75,09%). pada dataset lokal, pengujian tanpa adaptasi menghasilkan map@0,5 sebesar 6,07%, menunjukkan domain shift yang kuat. setelah fine-tuning dengan pembagian 70/15/15, kinerja meningkat menjadi 92,64% (citra asli) dan 93,74% (citra ditingkatkan) dengan recall 0,92. sistem berhasil dijalankan sepenuhnya offline pada jetson orin nano dan menampilkan keluaran deteksi melalui streamlit. kata kunci: deteksi ikan bawah air, peningkatan citra, clahe, yolov3, domain shift, fine-tuning, jetson orin nano, streamlit.

Underwater fish detection is essential for marine ecosystem monitoring, yet detection performance is often degraded by low contrast, bluish–green color dominance, turbidity, and non-uniform illumination. This paper proposes an end-to-end underwater fish detection pipeline that integrates CLAHE-based image enhancement with YOLOv3 object detection and validates its deployment on an NVIDIA Jetson Orin Nano edge-AI device using a Streamlit interface. Four CLAHE-based enhancement variants are evaluated: CLAHE, CLAHE with Unsharp Masking (USM), CLAHE with High-Frequency Emphasis Filtering (HEF), and CLAHE combined with Percentile Blending. Image quality is assessed on a 10% subset of the DeepFish dataset using LOE, UIQM, and UCIQE metrics. The results show that CLAHE with Percentile Blending provides the best trade-off, achieving the lowest LOE (212.021) and the highest UIQM (3.949) and UCIQE (0.508). Detection performance is evaluated on two domains: the DeepFish dataset and a locally collected dataset from Sabang, Aceh (Indonesia) to assess domain shift. On DeepFish, image enhancement yields an incremental improvement from 96.15% to 97.05% mAP@0.5, accompanied by a reduction in false negatives (198 to 137) and an increase in mean IoU (74.39% to 75.09%). In contrast, direct testing on the local dataset without adaptation results in a severe performance drop (6.07% mAP@0.5). After fine-tuning on the local dataset using a 70/15/15 train/validation/test split, performance improves substantially to 92.64% mAP@0.5 on original inputs and further to 93.74% on enhanced inputs, with a recall of 0.92. The proposed system operates fully offline on the Jetson Orin Nano and provides detection outputs via the Streamlit interface. Keyword: underwater fish detection, image enhancement, CLAHE, YOLOv3, domain shift, fine-tuning, Jetson Orin Nano, Streamlit.