Awaludin, Shazi (2023) PENERAPAN INSTANCE SEGMENTATION MENGGUNAKAN MASK R-CNN UNTUK MENGURANGI BIAS PADA HASIL KLASIFIKASI PENYAKIT DAUN TANAMAN KENTANG DAN ANGGUR. Other thesis, UPN "Veteran" Yogyajarta.
Text
ABSTRAK.pdf Download (148kB) |
|
Text
COVER.pdf Download (174kB) |
|
Text
DAFTAR ISI.pdf Download (143kB) |
|
Text
DAFTAR PUSTAKA.pdf Download (172kB) |
|
Text
PENGESAHAN PEMBIMBING.pdf Download (2MB) |
|
Text
PENGESAHAN PENGUJI.pdf Download (2MB) |
|
Text
SKRIPSI FULL - 123190123 - SHAZI AWALUDIN.pdf Restricted to Repository staff only Download (13MB) |
Abstract
vii
ABSTRAK
Pendeteksian penyakit tanaman sebelumnya dilakukan menggunakan image
classification yang bekerja dengan cara mengklasifikasi gambar terhadap label tertentu.
Hal tersebut cukup berhasil dilakukan pada model dengan arsitektur berbasis CNN.
Sayangnya sistem dengan teknik tersebut memiliki keterbatasan, yaitu model hanya
mampu untuk mendeteksi dengan baik apabila gambar yang dideteksi adalah gambar yang
semirip mungkin dengan data latih. Hal tersebut menjadi bias dan disebut sebagai
contextual bias.
Penelitian ini menggunakan instance segmentation pada dataset PlantVillage untuk
mendeteksi penyakit tanaman kentang dan anggur. Kedua tanaman tersebut dipilih karena
memiliki nilai dan dampak ekonomi yang tinggi bagi Indonesia. Penggunaan instance
segmentetion bertujuan mengurangi bias yang terjadi pada sistem pendeteksian penyakit
tanaman dan memungkinkan sistem untuk mendeteksi secara spesifik daun mana yang
terserang penyakit dan lokasinya secara akurat pada citra.
Pelatihan model dilakukan pada berbagai iterasi mulai dari 675 hingga 67500.
Pengujian dilakukan menggunakan 90 citra dimana 75 diantaranya adalah citra dengan dua
atau lebih objek daun, serta 15 diantaranya adalah citra dengan dua objek atau lebih selain
daun. Didapatkan bahwa iterasi 33750 menghasilkan performa terbaik diantara yang
lainnya dengan perolehan AP@[.50:.05:.95] pada bbox sebesar 79.36. dan mask sebesar
78.80.
Kata Kunci: penyakit tanaman, kentang, anggur, instance segmentation
viii
ABSTRACT
Plant disease detection was previously done using image classification which works
by classifying images against certain labels. This is quite successfully done on models with
CNN-based architecture. Unfortunately, the system with this technique has limitations,
where the model is only able to detect well if the detected image is as similar as possible to
the training data. This becomes a bias and is referred to as contextual bias.
This research uses instance segmentation on the PlantVillage dataset to detect potato
and grape diseases. Both crops were chosen because they have high economic value and
impact for Indonesia. The use of instance segmentation aims to reduce the bias that occurs
in plant disease detection systems and allows the system to detect specifically which leaves
are affected by disease and their location accurately in the image.
Model training was performed at various iterations ranging from 675 to 67500. The
training results show that the performance improvement is not so significant starting at
iteration 30000. The tests were conducted using 90 images of which 75 were images with
two or more leaf objects, and 15 were images with two or more objects besides leaves. It is
found that iteration 33750 produces the best performance among others with the
acquisition of AP@[.50:.05:.95] on bbox of 79.36 and mask of 78.80.
Keywords: plant diseases, potato, grape, instance segmentation
Item Type: | Thesis (Other) |
---|---|
Uncontrolled Keywords: | plant diseases, potato, grape, instance segmentation |
Subjects: | Z Bibliography. Library Science. Information Resources > ZA Information resources |
Divisions: | Faculty of Engineering, Science and Mathematics > School of Engineering Sciences |
Depositing User: | Eko Yuli |
Date Deposited: | 07 Sep 2023 05:06 |
Last Modified: | 07 Sep 2023 05:06 |
URI: | http://eprints.upnyk.ac.id/id/eprint/37405 |
Actions (login required)
View Item |