Even after multiple training attempts, my model plateaus at ~93% accuracy, the MSE does not improve, and the confusion matrix shows frequent misclassifications (especially diseased seeds being predicted as healthy). I tried: Custom CNN and transfer learning (ResNet50, VGG16). Data augmentation (rotation, flip, zoom, shift). Dropout, Batch Normalization, and learning rate schedulers. Adam optimizer with categorical crossentropy. Ensured balanced classes and 80/20 train-validation split. I expected higher accuracy (like near to 95%), better confusion matrix results, and reduced MSE given the dataset size. Continue reading...
