Pengembangan Metode Restorasi Citra pada Hasil Klasifikasi Posisi Tubuh Kuda

Abstract

Kolik menjadi salah satu penyebab utama kematian dini pada kuda sehingga memerlukan deteksi dini diantaranya melalui pengawasan Closed-Circuit Television (CCTV) dan model klasifikasi posisi tubuh kuda. Model YOLOv8 yang sudah dikembangkan menghasilkan citra dengan resolusi menurun akibat augmentasi, sehingga diperlukan restorasi citra agar hasil klasifikasi menyerupai citra asli. Penelitian ini bertujuan mengembangkan dan mengevaluasi metode restorasi citra menggunakan teknik interpolasi, serta menentukan kerangka kerja yang sesuai untuk mengembalikan resolusi citra sambil mempertahankan posisi bounding box. Data berupa 600 citra CCTV kandang kuda di Equestrian Park IPB dibagi menjadi dataset_1 (tanpa padding) dan dataset_2 (dengan padding). Empat model YOLOv8 dibangun dan dilatih dengan komposisi dataset berbeda, kemudian citra hasil klasifikasi di-upscale ke resolusi 1280×720 piksel menggunakan interpolasi Lanczos4, Cubic, Linear, dan Nearest, disertai penyesuaian bounding box. Evaluasi menggunakan metrik Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), dan Intersection over Union (IoU). Hasil menunjukkan model yang dilatih dengan dataset_2 memberikan kinerja klasifikasi terbaik dengan accuracy 0,9966 (dataset_1) dan 0,9983 (dataset_2), serta IoU tertinggi pada model 4 dengan 0,961. Metode Lanczos4 memberikan hasil restorasi terbaik pada dataset_1 dengan nilai SSIM sebesar 0,899025 dan nilai PSNR sebesar 27,974437, meski selisihnya terhadap metode lain tidak signifikan.

Colic is one of the leading causes of premature death in horses, necessitating early detection through CCTV (Closed-Circuit Television) surveillance and a horse body-posture classification model. The developed YOLOv8 model produces images with reduced resolution due to augmentation, thereby requiring image restoration so that the classification results resemble the original images. This study aims to develop and evaluate an image restoration method using interpolation techniques, and to determine an appropriate framework for restoring image resolution while preserving bounding box positions. The data comprised 600 CCTV images of horse stables at IPB Equestrian Park, divided into dataset_1 (without padding) and dataset_2 (with padding). Four YOLOv8 models were built and trained using different dataset compositions. The resulting classification images were then upscaled to a resolution of 1280×720 pixels using Lanczos4, Cubic, Linear, and Nearest interpolation, accompanied by bounding box adjustment. Results indicate that model that trained with dataset_2 achieved the best classification performance, with an accuracy of 0.9966 (dataset_1) and 0.9983 (dataset_2), and the highest IoU of model 4 is 0.961. The Lanczos4 method yielded the best restoration performance on dataset_1, with an SSIM of 0.899025 and a PSNR of 27.974437, although the difference from the other methods was not statistically significant.