PRS2Net: an efficient intelligent carrot detection model via filter pruning and attention mechanisms.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Carrots are nutrient-rich, vital for human health.
• Existing deep learning models for carrot quality assessment suffer from high computational costs and parameter redundancy.
• Efficient automated quality inspection is crucial for the agricultural and food industries.

Purpose of the Study:

• To develop a lightweight and efficient deep learning network for automated carrot quality inspection.
• To address the limitations of high computational costs and redundant parameters in current models.
• To enhance the speed and accuracy of carrot quality evaluation.

Main Methods:

• Compared four deep learning networks: GoogLeNet, MobileNet-v2, ResNet18, and ResNet50.
• Selected ResNet18 as the base for a novel lightweight network, PRS2Net.
• Applied pruning via first-order Taylor expansion and incorporated an attention mechanism to optimize ResNet18.

Main Results:

• PRS2Net reduced learnable parameters from 11,173,764 to 444,152.
• Achieved 97.25% accuracy on the validation set.
• Decreased training time by approximately 53.15% compared to the original ResNet18.

Conclusions:

• PRS2Net offers a significant improvement in the speed and efficiency of carrot quality evaluation.
• The developed lightweight network provides a practical, resource-efficient solution for automated agricultural systems.
• This advancement can lead to reduced operational costs and enhanced scalability in food industry applications.