Mitigating Spatial Scale Loss in CNN-Based Fine-Grained Image Classification: Application to Date Fruit Grading
Abstract
Accurate classification of date fruit varieties and size grades is critical for automated grading and post-harvest quality assessment. However, conventional image preprocessing techniques based on uniform resizing often distort size-dependent visual cues, leading to misclassification among size levels within the same variety. To address this limitation, this study proposes a size-preserving rescaling strategy for deep learning–based date fruit classification. Experiments are conducted on a curated dataset comprising 5,836 images distributed across 12 classes, representing four date varieties (Aseel, Dandhi, Karblain, and Kupro), each categorized into three size levels: large, medium, and small. Five convolutional neural network architectures—MobileNetV3, DenseNet121, InceptionV3, ResNet101, and VGG16—are evaluated using identical training, validation, and test splits under a supervised learning framework. When standard resized inputs are used, the highest classification accuracy achieved is 82.18%, with macro F1-scores close to 0.82. In contrast, incorporating the proposed size-preserving rescaling approach leads to substantial performance improvements across all models. The best results are obtained with ResNet101, achieving an accuracy of 94.44%, a macro precision of 0.9476, and a macro F1-score of 0.9446, followed closely by DenseNet121 with 94.32% accuracy. These findings demonstrate that preserving size information during preprocessing significantly enhances class separability and reduces size-level confusion, making the proposed approach well suited for practical date fruit grading systems.