<![CDATA[The SqueezeNet architecture is known to be effective but possesses a considerable number of parameters, which can be optimized using pruning—a compression technique that significantly reduces model parameters without sacrificing accuracy. This research aims to apply the L2-Norm based pruning method to the SqueezeNet architecture and compare its performance (accuracy and efficiency) against the default SqueezeNet model for classifying four classes of oil palm leaf diseases. The study used a primary dataset of 4,000 images, divided into training (70%), validation (20%), and testing (10%) sets. The SqueezeNet architecture was pruned using L2-Norm structured pruning with a uniform distribution at rates from 10% to 50%, followed by fine-tuning. The results show that the default SqueezeNet achieved 97.50% accuracy with 724,548 parameters. Significantly, a 10% pruning rate actually increased the accuracy to a high of 99.25% while simultaneously reducing the parameters to 579,036. Overly aggressive pruning, such as 40%, drastically decreased accuracy to 93.25%. It is concluded that the 10% pruning rate is the most optimal, proving that this method not only makes SqueezeNet lighter but also more effective. This 10% pruned model is highly suitable for application implementation due to its enhanced efficiency. Future research is recommended to validate these findings using a more diverse dataset and to test the model on actual edge devices.]]>
Copyrights © 2026
