Vpliv kvalitete vhodne slike na robustnost in natančnost zaznavanja objektov

Zalaznik, Jure (author), Logar, Vito (mentor)

Zaznavanje objektov je ključna naloga, ki se uporablja v številnih industrijah od avtonomne vožnje do industrijske varnosti, kjer omogoča avtomatski nadzor nošenja osebne varovalne opreme z namenom zagotavljanja večje varnosti pri delu. Modeli, kot je YOLO11 dosegajo visoko natančnost pri zaznavanju objektov na kakovostnih slikah, vendar se zmogljivost bistveno zmanjša, ko so vhodne slike izpostavljene različnim vrstam degradacije. Namen tega dela je sistematično preučiti, kako posamezne vrste slikovnih degradacij vplivajo na zmogljivost modela YOLO11 in na podlagi rezultatov identificirati kritične meje, pri katerih zaznavanje postane nezanesljivo. V teoretičnem delu so predstavljene osnove zaznavanja objektov, arhitektura YOLO11 ter ključne evalvacijske metrike: natančnost, priklic, presek nad unijo (IoU), F1-mera in srednja povprečna natančnost (mAP). Podatkovna množica je bila zasnovana s sedmimi razredi: pet razredov prisotne opreme (čelada, rokavice, obutev, varnostni brezrokavnik, človek) in dva razreda za zaznavanje kršitev (brez čelade, brez rokavic). Takšna razdelitev omogoča ne le zaznavanje varovalne opreme, temveč tudi neposredno identifikacijo kršitev varnostnih pravil. Model YOLO11 je bil treniran na čistih slikah in evalviran na čistih ter umetno degradiranih testnih slikah. Uspešnost modela na čistih slikah je bila ovrednotena z metrikami natančnosti, priklica, F1-mere in mAP, medtem ko je bil vpliv petih vrst degradacij (Gaussov beli šum, impulzni šum, gibalna zameglitev, kompresija JPEG ter sprememba svetlosti in kontrasta) analiziran na podlagi povprečne stopnje zaupanja. Na čistih slikah je model dosegel mAP@0,5 = 0,955, F1-mero 0,92 ter povprečno stopnjo zaupanja 0,63. Analiza robustnosti je pokazala, da je impulzni šum najbolj kritična degradacija, saj pri višjih stopnjah povzroči popolno odpoved zaznavanja. Močan negativen vpliv ima tudi gibalna zameglitev, medtem ko Gaussov beli šum povzroča postopno zmanjševanje zaupanja detekcij z naraščajočo stopnjo degradacije. JPEG kompresija se je izkazala za najmanj škodljivo degradacijo, sprememba svetlosti in kontrasta pa ima na delovanje modela le majhen vpliv. Rezultati potrjujejo primernost modela YOLO11 za avtomatski nadzor nošenja osebne varovalne opreme v industrijskih okoljih. Delo prispeva k boljšemu razumevanju vpliva degradacij na delovanje modela YOLO11, ter poudarja pomen ustrezne pred obdelave slik za zagotavljanje zanesljivega delovanja v realnih okoljih.

• Language: Slovenian
• Keywords: zaznavanje objektov, YOLO11, degradacije slike, Google Colab, osebna varovalna oprema, računalniški vid
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2026
• PID: 20.500.12556/RUL-186949
• Publication date in RUL: 07.09.2026

Secondary language

• Language: English
• Title: Impact of Input Image Quality on the Robustness and Accuracy of Object Detection

Abstract

Object detection is an important task in computer vision and is widely used in applications ranging from autonomous driving to industrial safety, where it enables automatic monitoring of personal protective equipment compliance and improves workplace safety. Models like YOLO11 perform very well on clear, high-quality images, but their performance decreases when the images are degraded. The goal of this thesis is to examine how different types of image degradation affect the YOLO11 model and to determine the points at which object detection becomes unreliable. The theoretical part explains the fundamentals of object detection, the YOLO11 model, and the main evaluation metrics, including precision, recall, Intersection over Union (IoU), F1-score, and mean Average Precision (mAP). A custom dataset with seven classes was created: five object classes (helmet, gloves, safety footwear, safety vest, person) and two classes representing safety violations (without helmet and without gloves). This setup allows not only the detection of protective equipment but also the direct identification of safety rule violations. The YOLO11 model was trained on clean images and tested on both clean and artificially degraded images. The model’s performance on clean images was evaluated using precision, recall, F1 score, and mAP, while the impact of five types of image degradation: Gaussian white noise, impulse noise, motion blur, JPEG compression and changes in brightness and contrast was analysed using the average confidence of detections. On clean images, the model achieved an mAP@0.5 of 0.955, an F1 score of 0.92, and an average confidence score of 0.63. The results showed that impulse noise had the strongest negative effect and could completely prevent the model from detecting objects at higher levels. Motion blur also reduced performance significantly, while Gaussian white noise caused a gradual decrease in detection confidence. JPEG compression had the smallest impact, changes in brightness and contrast only slightly affected the results. The results show that the YOLO11 model is suitable for automatically monitoring the use of personal protective equipment in industrial environments. This thesis also improves understanding of how image quality affects object detection and highlights the importance of proper image preprocessing for reliable performance in real-world situations.

• Keywords: object detection, YOLO11, image degradation, personal protective equipment, computer vision, industrial safety