Primerjava delovanja modelov strojnega vida

PAVLI, TIM (author), Meža, Marko (mentor)

Danes je naloga strojnega vida raziskovanje načinov, kako narediti računalnike bolj inteligentne in sposobne obdelave ter razumevanja senzoričnih informacij. Poleg tega je v številnih panogah naraščalo povpraševanje po avtomatiziranih sistemih, ki bi lahko izvajali naloge prepoznave in zaznave objektov, sledenja objektov ter analiziranja slik. Razvilo se je več modelov strojnega vida na podlagi različnih tehnik in metod, ki vključujejo globoko učenje in nevronske mreže. To jim omogoči različne sposobnosti zaznavanja glede na hitrost, natančnost in računsko zahtevnost. V nalogi so predstavljene tudi vse tematike, potrebne za razumevanje ozadja delovanja teh modelov. Cilj diplomske naloge je raziskati tehnike za strojno zaznavo in prepoznavo objektov. Za podrobnejšo primerjavo so izbrani trije algoritmi za podrobnejšo primerjavo. Ti algoritmi so: SSD, Faster RCNN in YOLO, na podlagi katerih so zgrajeni modeli za detekcijo objektov. Da lahko modeli opravljajo svojo nalogo, morajo biti naučeni na pripravljeni učni podatkovni zbirki. Na koncu se naredi primerjava glede na kriterije, kot so: gotovost zaznave, povprečna natančnost, hitrost obdelave, FPS, zasedenost računalniških virov in ocena zahtevnosti uporabe modelov. Predpogoj za učenje in zaganjanje modelov je dovolj velik pomnilnik, močan procesor in grafična kartica. Naloga opisuje postopke za pripravo lastnih testnih podatkovnih zbirk in izdelavo modelov. En primer učenja je izveden na modelu YOLOv8 in drugi s pomočjo Teachable Machine. Programi in kode za zagon modelov so napisane v programskem okolju Python, ki ga pripravimo z odprtokodno distribucijo Anaconda. Potrebni knjižnici za delovanje modela YOLOv8 sta Ultralytics in PyTorch. Naslednja modela sta SSD MobileNet FPN in Faster RCNN ResNet50, ki potrebujeta knjižnico TensorFlow in knjižnico openCV. Da dobimo učinkovito primerjavo, so ti modeli naučeni na kakovostni podatkovni zbirki COCO. Naša primerjava je pokazala, da je bil najbolj učinkovit model YOLOv8, saj omogoča zaznavo in prepoznavo objektov v realnem času z dobro natančnostjo. Hkrati tudi ne potrebuje zelo močnega grafičnega procesorja in je uporabniku prijazen za implementacijo. Modela SSD MobileNet FPN in Faster RCNN ResNet50 sta medtem programersko zahtevnejša za uporabo. Model Faster RCNN je dosegel tudi največje gotovosti zaznave, vendar je njegov odziv počasen in porabi največ računalniških virov. Samo model SSD ni dosegel pričakovanih rezultatov, saj se pri slabi natančnosti tudi ni dovolj hitro odzval. Uporaba pravega modela je na koncu odvisna od potrebe uporabnika za reševanje določenega problema.

• Language: Slovenian
• Keywords: strojni vid, globoko učenje, konvolucijske nevronske mreže, detekcija objektov, primerjava modelov, SSD, Faster RCNN, YOLO
• Work type: Bachelor thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2023
• PID: 20.500.12556/RUL-149882
• COBISS.SI-ID: 165035011
• Publication date in RUL: 11.09.2023

Secondary language

• Language: English
• Title: Comparative analysis of computer vision models performance

Abstract

The task of machine vision today is to explore ways to make computers more intelligent and capable of processing and understanding sensory information. In addition, there was a growing demand in many industries for automated systems that could perform the tasks of object recognition and detection, object tracking, and image analysis. Several machine vision models have been developed based on various techniques and methods, including deep learning and neural networks. This enables them to have different cognitive abilities in terms of speed, accuracy and computational complexity. The assignment also presents all the topics necessary to understand the background of the operation of these models. The diploma thesis aims to investigate a technique for machine detection and recognition of objects, from which three algorithms are selected for a more detailed comparison. The selected algorithms are SSD, Faster RCNN and YOLO, based on which models for object detection are built. To perform their task, models must be trained on a prepared training database. In the end a comparison is made according to criteria such as certainty of detection, average accuracy, processing speed, FPS, use of computer resources, and assessment of the complexity of using the models. A prerequisite for learning and running models is a sufficiently large memory, a powerful processor and a graphic card. The task describes the procedures for preparing own test data sets and creating models. One learning example is performed on the YOLOv8 model and the other using the Teachable Machine. Programs and codes for running the models are written in the Python programming environment, prepared by the open-source Anaconda distribution. The libraries required to run the YOLOv8 model are Ultralytics and PyTorch. The next models are SSD MobileNet FPN and Faster RCNN ResNet50, which requires the TensorFlow library and the openCV library. To obtain an effective comparison, all these models are trained on the high-quality COCO database. Our comparison showed that the YOLOv8 model is the most effective, because it enables real-time object detection and recognition with good accuracy. At the same time, it does not require a very powerful graphics processor and is user-friendly to implement. Meanwhile, the SSD MobileNet FPN and Faster RCNN ResNet50 models are more demanding to use as programmers. The Faster RCNN model also achieved the highest detection certainty, but its response is slow and consumes the most computing resources. Only the SSD model did not achieve the expected results, due to its poor accuracy and lack of quick response. The use of the right model in the end depends on the user's need to solve a particular problem.

• Keywords: computer vision, deep learning, convolutional neural networks, object detecton, model comparison, SSD, Faster RCNN, YOLO