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

Carrots, rich in essential nutrients, play a crucial role in supporting human health. Current deep learning networks for carrot quality inspection are constrained by redundant parameters and high computational costs. To address these issues, this paper introduces a lightweight network, PRS2Net, based on ResNet18 selected after comparing four networks (GoogLeNet, MobileNet-v2, ResNet18, ResNet50). ResNet18 was pruned using first-order Taylor expansion to reduce redundancy and enhanced with an attention mechanism to focus on critical features.

PRS2Net achieved high efficiency with learnable parameters reduced from 11 173 764 to 444 152, while maintaining 97.25% accuracy on the validation set. Training time was cut by about 53.15% compared to ResNet18, significantly speeding up carrot quality inspection.

This approach enhances the speed and efficiency of carrot quality evaluation, offering a practical, resource-efficient solution for real-world applications in agriculture and food industries, potentially reducing operational costs and improving scalability for automated systems. © 2025 Society of Chemical Industry.