20.500.12556/RUL-171939 Segmentacija vinskih trt iz slikovnih podatkov s temeljnimi modeli Vineyard segmentation from image data with foundation models V vinogradništvu je treba natančno analizirati stanje rastlin in redno spremljati njihovo rast. Tradicionalne metode, ki temeljijo na ročnih pregledih vsake rastline, so zamudne in pogosto neučinkovite. Zato je nujno razviti avtomatiziran sistem, ki omogoča hitro, natančno in objektivno analizo velikih površin. Cilj diplomske naloge je razviti avtomatiziran sistem za zaznavanje oz. segmentacijo trt, ki bo zagotavljal zanesljive in natančne rezultate. Tako bomo omogočili učinkovito analizo obsežnih vinogradnih površin brez potrebe po ročnih pregledih vsake posamezne rastline. Pri reševanju problema avtomatizirane segmentacije trt uporabljamo metode računalniškega vida in globokega učenja, pri čemer je glavno orodje model SAM (Segment Anything Model), natančneje njegova najnovejša različica SAM 2. Ključni del metodologije vključuje predhodno označevanje slik, učenje modela in optimizacijo rezultatov. Pri tem uporabljamo knjižnice, kot so PyTorch, NumPy in OpenCV, ki omogočajo prilagodljivo obdelavo podatkov in vizualizacijo rezultatov. Za ovrednotenje modela uporabljamo metrike, kot sta IoU (presek nad unijo) in ocena F1, s katerima ocenjujemo uspešnost segmentacije glede na ročno označene slike. Končne rezultate primerjamo z osnovnim modelom SAM 2. In viticulture, it is necessary to accurately analyze the condition of plants and regularly monitor their growth. Traditional methods based on manual inspections of each plant are time-consuming and often inefficient. Therefore, it is imperative to develop an automated system that enables fast, accurate and objective analysis of large areas. The goal of the thesis is to develop an automated vineyard detection system that will provide reliable and accurate results. This will enable efficient analysis of large vineyard areas without the need for manual inspections of each individual plant. We approach the vineyard detection problems through a segmentation task for which we use computer vision and deep learning methods, with the main tool being the SAM (Segment Anything Model) model, particularly, its latest version SAM 2. The key part of the methodology includes pre-labeling of images, model training and optimization of results. We use libraries such as PyTorch, NumPy and OpenCV, which enable flexible data processing and visualization of results. To validate the model, we use metrics such as IoU (Intersection over Union) and F1-score, which evaluate the success of segmentation based on manually labeled images. The final results are compared with the base SAM 2 model. globoko učenje segmentacija vinograd nevronske mreže doučitev SAM SAM 2 deep learning segmentation vineyard neural networks fine-tuning SAM SAM 2 true false false Slovenski jezik Angleški jezik Diplomsko delo/naloga 2025-09-04 10:00:00 2025-09-04 10:00:09 2025-09-28 03:51:17 0000-00-00 00:00:00 2025 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 62906 248583427 Saranovic_Amar_-_Segmentacija_vinskih_trt_iz_slikovnih_podatkov_s_temeljnimi_modeli.pdf Saranovic_Amar_-_Segmentacija_vinskih_trt_iz_slikovnih_podatkov_s_temeljnimi_modeli.pdf 1 EF95F499CA3E71047323B55CABD5DCA2 19be5e0c1818093f17442c7029066eb75fb759c96a95b02a26e0ec898c566f11 093b6aea-8965-11f0-9328-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=215523 Fakulteta za elektrotehniko 0 0 0