Denarda, Alessandro Rocco (2024) LABEL-EFFICIENT WEAKLY SUPERVISED FLOWER AND FRUIT DETECTION FOR OPTIMIZED CROP MANAGEMENT. [Tesi di dottorato]

Documento PDF Tesi_Dottorato_Alessandro_Rocco_Denarda_37Ciclo_AI_Agrifood_e_Ambiente.pdf Visibile a [TBR] Amministratori dell'archivio Download (91MB) | Richiedi una copia

Abstract

The rapid advancements of Artificial Intelligence technologies have opened up new possibilities for several kinds of agricultural practices and applications. Among these, the use of Computer Vision systems for crop monitoring and management has emerged as a key focus area, offering significant benefits in automating agricultural processes. Several tasks have seen major improvements through the use of Computer Vision, including disease and pest monitoring, plant vigor assessment, yield estimation via flower and fruit counting, and the optimization of irrigation and fertilization practices. Flower and fruit detection has garnered particular attention, as it plays a crucial role in estimating agricultural yield. Fruit detection and counting represent one of the most important steps toward yield estimation and a well-known practice for farmers, on which they base the management of the harvesting, storage, and distribution phases of products. Flowering assessment in tree-fruit orchards is also crucial for harvest management in the context of estimation of both potential yield and the time when fruits will reach harvest maturity. These practices, which were previously performed only by human operators, are currently being redesigned through the employment of Computer Vision techniques. However, despite the impressive results achieved by these systems, they rely on the availability of large image datasets, which is still limited if compared to the great number of crop typologies. One of the main responses to this challenge is represented by data sharing, which has recently gained widespread recognition as a key factor in saving resources required for data collection and annotation, and in facilitating the benchmark of Computer Vision frameworks across research groups. Additionally, great interest has recently been devoted to Weakly Supervised frameworks, which aim to reduce the effort required for dataset annotation by utilizing simple image-level labels during trainings. The current study aimed to contribute to the design and implementation of easily applicable systems for flower and fruit detection and counting using a Weakly Supervised approach. Three distinct frameworks were developed, leveraging the capabilities of Encoder-Decoder Convolutional Neural Networks trained with presence/absence annotations to provide accurate activation maps for instance detection. The frameworks were tested in various scenarios, including mango orchards at different growth stages and olive orchards. Additionally, two novel datasets focusing on saffron flower crops and olive orchards were introduced for validating and testing Computer Vision-based object detection methodologies.

Downloads

Actions (login required)

Modifica documento