Detekcija kapljica kiše na zaštitnom staklu između scene i kamere

master's thesis

• Title (english): Detection of raindrops on the protective glass between the scene and the camera
• Author: Ivana Barišić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-13, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Raspoznavanje slika važno je područje računalnog vida s mnogim zanimljivim primjenama. Jedna od takvih primjena je i detekcija smanjene vidljivosti uslijed kapljica na zaštitnom staklu kamere. Ovaj rad proučava rješavanje tog problema prikladno oblikovanim konvolucijskim modelima, točnije korištenjem modela s ResNet-18 arhitekturom. Raznim postupcima, poput učenja na isječcima, uvođenjem dilatirane konvolucije te smanjenjem koraka u prvom sloju sažimanja, pokušava se dobiti što bolja sposobnost generalizacije. Fokus je stavljen na prepoznavanja sitnih detalja slike i očuvanje originalne ImageNet semantike mreže. U sklopu rada korištena su dva skupa podataka od kojih su oblikovani podskupovi za učenje, validaciju i testiranje. Kreiran je klasifikacijski algoritam strojnog učenja te su vrednovani naučeni modeli i prikazane su njihove metrike kao i vizualizirana logika zaključivanja modela pomoću gradijenata ulaza.

Abstract (english)

Image recognition is an important area of computer vision with many interesting applications such as the detection of reduced visibility due to drops on the protective glass of the camera. This paper studies the solution to this problem with appropriately designed convolutional models, i.e. using ResNet-18 models. The best possible model was deducted by a combination of various procedures, such as learning on clips, introducing dilated convolution and reducing stride in the first pooling layer. The focus was on recognizing the fine details of the image and preserving the original ImageNet semantics of the network. Subsets for learning, validation and testing were formed from two data sets which were used in the paper. The classification algorithm of machine learning was created and the learned models were evaluated, along with their metrics, as well as the visualized logic of the model inference using input gradients to do so.

• Keywords: detekcija kapljica; dilatirana konvolucija; ResNet-18; vizualizacija mapa istaknutih područja
• Keywords (english): raindrop detection; dilated convolution; ResNet-18; saliency maps visualization
• Language: croatian
• URN:NBN: urn:nbn:hr:168:653967
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 08:24:22