20.500.12556/RUL-188868 Prepoznavanje bolezni rastlin ob omejeni količini podatkov Identifying plant diseases with limited data Bolezni rastlin povzročajo veliko škodo v kmetijski industriji. Strokovnjaki ocenjujejo, da zaradi bolezni vsako leto izgubimo več kot 30% pridelka. Poleg manjše količine pridelka bolezni poslabšajo tudi njegovo kakovost. Da bi ustavili širjenje bolezni, je pomembno njihovo zgodnje odkrivanje. Pri tem si danes lahko pomagamo s tehnologijo. Modeli globokega učenja za prepoznavanje teh bolezni kažejo dobre rezultate, a za učenje pogosto zahtevajo veliko količino slik, ki pa jih pogosto nimamo. V magistrskem delu smo razvili sistem, ki za učenje potrebuje relativno majhne količine podatkov. Naš algoritem na sliki najprej poišče in izolira posamezen list ali plod. Zatem ga razdeli na manjše dele, iz njih izlušči ključne informacije in na podlagi teh rastlino umesti v določen razred (npr. »zdrav« ali »nezdrav«). Eksperimenti s slikami guave, pšenice in kave so pokazali, da naš pristop pri uporabi zgolj 1 do 10 učnih slik prepozna objekte zanesljiveje kot primerljivi modeli. Plant diseases are causing a lot of damage within the agriculture industry. Experts have estimated that because of diseases, around 30% of food crops are lost. Besides the lower quantities of produced crops, diseases also lower their quality. In order to stop the spread of diseases among the crops, it is important to discover them early. Today we can use technology to assist us. Deep learning models for detecting said diseases are showing good results, but they often require big amounts of images for training, which we often do not have. In our master's thesis we developed a system that requires relatively low amounts of data for training. Our algorithm first locates and segments a certain leaf or fruit. After that it divides it into smaller pieces and extracts key information from them, and based on that it later classifies the plant into a certain class (for example, »healthy« or »unhealthy«). Experiments with images of guava, wheat, and coffee showed that our approach, when using only 1 to 10 images, classifies images better than the comparing models. prepoznavanje bolezni rastlin omejene količine podatkov računalniški vid strojno učenje plant disease identification limited data computer vision machine learning true false false Slovenski jezik Angleški jezik Magistrsko delo/naloga 2026-09-29 17:05:03 2026-09-29 17:05:15 2026-09-30 04:29:23 0000-00-00 00:00:00 2026 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 1970-01-01 38434 Urankar_Jan_-_Prepoznavanje_bolezni_rastlin_ob_omejeni_kolicini_podatkov.pdf Urankar_Jan_-_Prepoznavanje_bolezni_rastlin_ob_omejeni_kolicini_podatkov.pdf 1 227DCFCA1311252801A503109E9BB45E 1f35fd469a3a0d6ef200fbbf93d18a0e53380459aa963bed262cae0d47ce46c3 14b21f40-bc17-11f1-8bc5-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=249246 Fakulteta za računalništvo in informatiko 0 0 0