Accurate and efficient quality classification is essential for ensuring consistency and market competitiveness in the orange industry. This study proposes a hybrid approach that combines autoencoder-based feature extraction with traditional machine learning models to improve the accuracy of orange quality assessment. We design and evaluate three different autoencoder (AE) architectures, and extract bottleneck features to represent compact and informative latent spaces. These learned features are then used to train various ML classifiers. Experiments on a public dataset show that the best performance is achieved by combining the third AE architecture with CatBoost, reaching an accuracy of 75.51%, which significantly outperforms the previous best result of 69.38% reported by the K-Nearest Neighbors (KNN) model. The results demonstrate that learned feature representations can effectively enhance the discriminative power of classifiers. This study highlights the potential of tailored deep feature extraction in agricultural product quality control and provides a promising direction for future research in representation learning for food quality assessment.