Electronic Theses and Dissertation

Penyakit daun cabai merupakan salah satu faktor yang menurunkan produktivitas tanaman sehingga diperlukan sistem identifikasi yang cepat dan akurat. Metode klasifikasi berbasis CNN telah banyak digunakan untuk mendeteksi penyakit daun cabai, namun performanya dipengaruhi kualitas citra. Variasi pencahayaan pada proses pengambilan citra, seperti perbedaan intensitas cahaya, bayangan, dan kontras tidak merata, dapat menyebabkan ketidakkonsistenan kualitas citra sehingga mengganggu ekstraksi fitur dan menurunkan akurasi klasifikasi. Penelitian bertujuan mengevaluasi efektivitas teknik normalisasi fotometrik dalam meningkatkan kualitas citra serta performa klasifikasi penyakit daun cabai menggunakan model CNN. Metode yang digunakan meliputi Homomorphic Filtering, High-Pass Filtering, High-Boost Filtering, dan Multi-Scale Retinex (MSR). Teknik tersebut digunakan untuk memperbaiki pencahayaan dan meningkatkan kualitas fitur citra sebelum proses klasifikasi. Dataset terdiri atas 1.631 citra daun cabai dengan 5 kelas, yaitu bercak, kuning, keriting, sehat, dan whitefly. Citra hasil normalisasi diuji menggunakan model VGG19, MobileNetV2, dan AlexNet. Evaluasi kualitas citra dilakukan menggunakan Visual Assessment, Mean Squared Error (MSE), dan Peak Signal-to-Noise Ratio (PSNR), sedangkan performa klasifikasi dianalisis menggunakan confusion matrix. Hasil penelitian menunjukkan bahwa teknik normalisasi fotometrik meningkatkan kualitas citra dan kemampuan model mengekstraksi fitur penyakit daun cabai, dengan High-Boost Filtering menghasilkan hasil visual terbaik berdasarkan rata-rata MSE 0,000821 dan PSNR 32,26 dB. Dari seluruh skenario pengujian, performa klasifikasi tertinggi diperoleh pada kombinasi High-Pass Filtering dan MobileNetV2 dengan akurasi 93,55%, precision 94,46%, recall 93,18%, dan F1-score 93,27%. Temuan ini mengindikasikan bahwa peningkatan kualitas citra melalui normalisasi fotometrik efektif mendukung klasifikasi penyakit daun cabai pada berbagai kondisi pencahayaan, meskipun kualitas citra terbaik tidak selalu menghasilkan performa klasifikasi tertinggi.

Chili leaf diseases are one of the factors that reduce crop productivity, making a fast and accurate identification system necessary. CNN-based classification methods have been widely used to detect chili leaf diseases, but their performance is influenced by image quality. Variations in lighting during image acquisition, such as differences in light intensity, shadows, and uneven contrast, can lead to inconsistent image quality, thereby disrupting feature extraction and reducing classification accuracy. This study aims to evaluate the effectiveness of photometric normalization techniques in improving image quality and the performance of chili leaf disease classification using a CNN model. The methods used include Homomorphic Filtering, High-Pass Filtering, High-Boost Filtering, and Multi-Scale Retinex (MSR). These techniques were used to correct lighting and improve image feature quality prior to the classification process. The dataset consists of 1,631 chili leaf images with 5 classes: spots, yellowing, curling, healthy, and whitefly. The normalized images were tested using the VGG19, MobileNetV2, and AlexNet models. Image quality was evaluated using Visual Assessment, Mean Squared Error (MSE), and Peak Signal-to-Noise Ratio (PSNR), while classification performance was analyzed using a confusion matrix. The results of the study show that photometric normalization techniques improve image quality and the model’s ability to extract features of chili leaf diseases, with High-Boost Filtering producing the best visual results based on an average MSE of 0.000821 and a PSNR of 32.26 dB. Across all testing scenarios, the highest classification performance was achieved with the combination of High-Pass Filtering and MobileNetV2, with an accuracy of 93.55%, precision of 94.46%, recall of 93.18%, and an F1-score of 93.27%. These findings indicate that improving image quality through photometric normalization effectively supports the classification of chili leaf diseases under various lighting conditions, although the best image quality does not always result in the highest classification performance.