Analisis Pengaruh Teknik Praproses Citra Pada Model Deteksi Kelainan Daun Tanaman Melon Menggunakan Algoritma YOLOv7

Abstract

Penelitian ini bertujuan untuk menganalisis dan menerapkan teknik praproses citra guna meningkatkan performa model deteksi kelainan daun melon menggunakan algoritma YOLOv7, menerapkan model terbaik pada Jetson Nano dan sistem berbasis website serta aplikasi yang terhubung dengan cloud firestore. Data yang digunakan adalah citra kelainan daun melon dengan total 521 citra, yang dibagi menjadi 90% data latih dan 10% data uji. Pengolahan data dilakukan melalui teknik augmentasi dan praproses yaitu Averaging Histogram Equalization (AVGHEQ), Brightness Preserving Dynamic Histogram Equalization (BPDFHE) dan Contrast Limited Adaptive Histogram Equalization (CLAHE) yang dibandingkan dengan set data asli. Hasil penelitian menunjukkan bahwa rata-rata nilai mAP model antara 58,6 – 66,3%, akurasi model antara 80,7 – 84,9%, dan waktu deteksi model antara 9,8 – 20 milidetik. Praproses citra mampu meningkatkan kinerja model dibandingkan dengan set data asli pada variabel mAP dan waktu deteksi, kecuali variabel akurasi. Ukuran kernel dan learning rate menunjukkan perbedaan signifikan dalam kinerja model. Ukuran kernel bernilai 3 lebih baik daripada 5 dan 7. Learning rate bernilai 0,001 lebih baik daripada 0,1 dan 0,01. Variasi fungsi aktivasi, pooling layer, batch size, dan momentum tidak menunjukkan berbeda nyata terhadap kinerja model. Model terbaik diperoleh pada pemodelan dengan menggunakan epoch maksimum dan augmentasi standar YOLOv7. Model dengan epoch maksimum mencapai mAP rata-rata 84,12%, akurasi 91,19%, dan waktu deteksi 4,55 milidetik. Di sisi lain, model dengan patience 300 memperoleh mAP rata-rata 81,57%, akurasi 92,23%, dan waktu deteksi 5,03 milidetik. Pemberian augmentasi pada konfigurasi awal YOLOv7 memberikan dampak yang signifikan terhadap performa model deteksi. Semakin banyak epoch yang diberikan maka hal tersebut akan menghasilkan model yang semakin kokoh sehingga dapat meningkatkan kinerja deteksi model. Implementasi model menggunakan mini komputer Jetson Nano meningkatkan sumber daya CPU sebesar 25% (dari 25% menjadi 50%) sebelum dan sesudah implementasi model dan penggunaan RAM meningkat sebesar 20% (dari 70% menjadi 90%). Waktu deteksi rata-rata pada Jetson Nano dengan model Ensemble sebesar 0,83 detik, sementara itu untuk model terbaik Fold 4 diperoleh sebesar 0,11 detik. Pengujian sistem berupa website dan aplikasi berjalan dengan baik.

This research aims to analyze and apply image preprocessing techniques to enhance the performance of a melon leaf anomaly detection model using the YOLOv7 algorithm, implement the best model on a Jetson Nano, and develop a web-based system and application connected to Cloud Firestore. The data used consists of 521 images of melon leaf anomalies, divided into 90% training data and 10% testing data. Data processing was conducted through augmentation techniques and preprocessing methods such as Averaging Histogram Equalization (AVGHEQ), Brightness Preserving Dynamic Histogram Equalization (BPDFHE) and Contrast Limited Adaptive Histogram Equalization (CLAHE), which were compared to the original set data. The results of the study showed that the average mAP value of the model ranged from 58.6% to 66.3%, the model accuracy ranged from 80.7% to 84.9%, and the model detection time ranged from 9.8 to 20 milliseconds. Image preprocessing was able to improve the model's performance compared to the original set data in terms of mAP and detection time, except for the accuracy variable. Kernel size and learning rate showed significant differences in model performance. Kernel size of 3 performs better than 5 and 7. Learning rate of 0.001 performs better than 0.1 and 0.01. Variations in activation functions, pooling layers, batch size, and momentum do not show significant differences in model performance. The best model was achieved by using the maximum epoch and default YOLOv7 augmentation. The model with the maximum epoch achieved an average mAP of 84.12%, an accuracy of 91.19%, and a detection time of 4.55 milisecond. On the other hand, the model with a patience of 300 achieved an average mAP of 81.57%, an accuracy of 92.23%, and a detection time of 5.03 milisecond. Augmentation in the initial configuration of YOLOv7 significantly impacted model performance. Increasing the number of epochs resulted in a more robust model, thus enhancing detection performance. Implementing the model using the Jetson Nano mini-computer increased CPU resource usage by 25% (from 25% to 50%) before and after model implementation, and RAM usage increased by 20% (from 70% to 90%). The average detection time on the Jetson Nano with the Ensemble model was 0.83 seconds, while for the best model Fold 4 it was obtained at 0.11 seconds. The system testing in the form of a website and application worked well.