Automated fruit recognition using deep learning techniques

Entity: Universitat Rovira i Virgili (URV)

Confidenciality: No

Education area(s): Enginyeria de la Seguretat Informàtica i Intel·ligència Artificial

APS: No

Title in different languages: Automated fruit recognition using deep learning techniques

Abstract: This thesis presents a comprehensive study on automated object recognition using deep learning approaches, with fruit recognition serving as a representative case study to demonstrate the broader applicability of computer vision methodologies. The research addresses the fundamental challenges in developing robust visual recognition systems by implementing and comparatively analyzing two distinct deep learning paradigms: transfer learning with ResNet50 for classification tasks and YOLOv11 for real-time object detection. The methodology encompasses both controlled and real-world scenarios to evaluate model performance across different environmental conditions. Using the Fruits- 360 dataset containing 201 categories with controlled backgrounds, the ResNet50 model achieved an exceptional accuracy of 98.62% through transfer learning techniques. In parallel, the YOLOv11 implementation, trained on a real-world dataset of 32 classes with natural backgrounds, demonstrated robust detection capabilities with 93.2% mAP@0.5 and real-time processing at 53.5 FPS. A critical contribution of this research is the systematic analysis of the domain gap between controlled laboratory conditions and real-world deployment scenarios. The findings reveal that while classification models excel on standardized datasets, their performance may degrade significantly when applied to natural environments, whereas detection models trained on diverse real-world data ensure broader practical applicability. To demonstrate the integration of both approaches, a unified graphical interface was developed, showcasing how complementary deep learning techniques can be combined for comprehensive visual recognition systems. While this work specifically targets agricultural automation applications such as sorting systems, quality control, and robotic harvesting, the methodological framework and comparative analysis provide valuable insights for any computer vision application requiring robust object recognition across varying environmental conditions. The research contributes to the broader field of automated visual recognition by establishing best practices for model selection, training strategies, and deployment considerations in real-world scenarios.

Subject: Visió artificial (Robòtica)

Academic year: 2024-2025

Language: en

Work's public defense date: 2025-06-12

Subject areas: Computer engineering

Student: Abderrahmane, Hammia

Department: Enginyeria Informàtica i Matemàtiques

Creation date in repository: 2026-03-13

TFM credits: 9

Keywords: Computer vision, Deep learning, Object recognition, ResNet50, YOLOv11, Transfer learning, Object detection, Domain adaptation, Agricultural automation

Title in original language: Automated fruit recognition using deep learning techniques

Access Rights: info:eu-repo/semantics/openAccess

Project director: Isern Alarcón, David