Tomato maturity recognition with convolutional transformers.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate tomato maturity classification is crucial for agricultural applications like harvesting and quality control.
• Existing methods may lack the precision needed for diverse real-world conditions.
• Developing robust automated systems is essential for modern agriculture.

Purpose of the Study:

• To propose a novel method for tomato maturity classification using a convolutional transformer.
• To introduce a new, diverse tomato image dataset (KUTomaData) for training deep learning models.
• To demonstrate the superiority of the proposed convolutional transformer over state-of-the-art methods.

Main Methods:

• Developed a hybrid convolutional transformer architecture combining CNNs and transformers.
• Created the KUTomaData dataset with ~700 images under varied lighting and perspectives.
• Evaluated the framework on KUTomaData and two public datasets (Laboro Tomato, Rob2Pheno Annotated Tomato).

Main Results:

• The convolutional transformer significantly outperformed state-of-the-art methods.
• Achieved mean average precision scores of 58.14% (KUTomaData), 65.42% (Laboro Tomato), and 66.39% (Rob2Pheno Annotated Tomato).
• The framework effectively handled cluttered and occluded tomato instances.

Conclusions:

• The proposed convolutional transformer is a highly effective approach for tomato maturity classification.
• The KUTomaData dataset provides a valuable resource for agricultural deep learning research.
• This work has the potential to enhance efficiency and accuracy in tomato production and processing.