Fruit freshness detection based on multi-task convolutional neural network.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Fruit freshness detection is crucial for agriculture, impacting automated harvesting and supply chain management.
• Computer vision techniques are increasingly vital for monitoring fruit quality and optimizing agricultural processes.

Purpose of the Study:

• To develop a deep learning model using multi-task learning (MTL) for enhanced fruit freshness detection.
• To investigate the effectiveness of MTL by comparing it with single-task learning (STL) for fruit freshness detection and classification.

Main Methods:

• Designed a multi-task learning model with a shared convolutional neural network (CNN) subnet and two task-specific fully connected (FC) heads.
• Optimized freshness detection (T1) and fruit type classification (T2) tasks in parallel.
• Conducted a comparative study using an open fruit image dataset, evaluating MTL against STL models.

Main Results:

• The MTL model achieved higher mean accuracies: 93.24% for freshness detection and 88.66% for classification, compared to STL's 92.50% and 87.22%.
• Statistical analysis confirmed significant performance differences between MTL and STL.
• Analysis of feature vectors revealed a strong correlation (average cosine similarity of 0.7) between the tasks, validating the MTL approach.

Conclusions:

• Multi-task learning effectively exploits correlations between related tasks to improve feature extraction.
• The proposed MTL approach offers a more efficient and accurate method for fruit freshness detection and classification.
• This methodology holds potential for extension to other domains involving interconnected tasks.