Autoencoder Applications for Enhancing Spice Classification Performance under Limited Dataset Conditions

Indonesia's rich assortment of spices holds immense value for their health benefits and culinary applications, yet widespread understanding of these spices among the populace remains limited. This research proposes leveraging autoencoder technology to address the complexities arising from the intricate shapes and colors of spices, compounded by the constraints of a limited dataset. Through meticulous refinement of hyperparameters such as learning rate, optimization function, input size, and epochs, the autoencoder endeavors to extract unique features crucial for accurate spice classification. Comparative assessments against alternative CNN architectures employing transfer learning underscore the autoencoder's capacity to attain comparable accuracy levels. Notably, the model demonstrates consistent performance devoid of overfitting or underfitting, yielding validation precision and recall rates nearing 0.97. These outcomes accentuate the autoencoder's efficacy in compressing image data and discerning pivotal features, thereby facilitating precise spice classification despite constraints posed by dataset limitations.