Classification Performance and Energy Efficiency of Pre-Trained Deep Learning Models for Chicken Meat Cuts

In recent years, advances in computer vision and deep learning have enabled automated classification and sorting systems across the food processing and agricultural industries. Automated identification of chicken meat cuts (e.g., breast, drumstick, thigh) is essential for streamlining automated packaging, inventory management, and processing lines. However, the high computational complexity and associated energy consumption of deep neural networks often restrict their deployment on resource-constrained edge platforms. This study investigates the trade-off between classification performance and energy consumption during inference among five pre-trained convolutional neural network (CNN) architectures—SqueezeNet, MobileNet-v2, EfficientNet-b0, GoogleNet, and ResNet-50—to classify five anatomical categories of chicken meat: breast, drumstick, thigh, whole-wing, and wing. Feature extraction was performed using these pre-trained networks, followed by classification with a fully connected layer. Classification performance was evaluated using accuracy, precision, sensitivity, specificity, and F1-score, while processor energy consumption during inference was measured using Intel Power Gadget. The experimental results demonstrate that EfficientNet-b0 achieves the highest overall classification accuracy (94.14%), while SqueezeNet exhibits the lowest measured energy consumption (2024.95 J) and the shortest inference latency (11.91 ms/image). Together, EfficientNet-b0 and SqueezeNet form the Pareto frontier for classification performance and energy efficiency during inference. SqueezeNet’s lightweight architecture enabled competitive accuracy while requiring significantly lower energy consumption than ResNet-50 and GoogleNet. The results indicate robustness to changes in specimen orientation and storage time under the evaluated conditions. While evaluated on a standardized processor baseline, these findings highlight the prospective viability of lightweight deep learning models for potential edge deployment in automated food sorting systems.