Healthy Harvests: A Comparative Look at Guava Disease Classification Using InceptionV3

Guava fruits often suffer from many diseases. This can harm fruit quality and fruit crop yield. Early identification is important for minimizing damage and ensuring fruit health. This study focuses on 3 different categories for classifying diseases. These are Anthracnose, Fruit flies, and Healthy fruit. The data set used in this study is collected from Mendeley Data. This dataset contains 473 original images of Guava. These images vary in size and format. The original dataset was resized to 256x256 pixels with RGB color mode for better consistency. After this, the Data augmentation process is applied to improve the dataset by generating variations of the original images. The augmented dataset consists of 3784 images using advanced preprocessing techniques. Two deep learning models were implemented to classify the images. The InceptionV3 model is well known for its advanced framework. These apply multiple convolutional filters for obtaining different features effectively. On the other hand, the ResNet50 model helps to train deeper networks by using residual learning. The InceptionV3 model achieved the impressive accuracy of 98.15%, and ResNet50got 94.46% accuracy. Data mixing methods such as CutMix and MixUp were applied to enhance the model’s robustness. The confusion matrix was used to evaluate the overall model performance of both InceptionV3 and Resnet50. Additionally, SHAP analysis is used to improve interpretability, which helps to find the significant parts of the image for the model prediction. This study purposes to highlight how advanced models enhan

💡 Research Summary

The paper presents a comparative study of two state‑of‑the‑art convolutional neural networks—InceptionV3 and ResNet50—for classifying guava fruit images into three categories: Anthracnose, Fruit fly infestation, and Healthy. The authors sourced 473 original guava images from Mendeley Data, resized them to a uniform 256 × 256 px RGB format, and then applied aggressive data augmentation to expand the dataset to 3,784 images. The augmentation pipeline incorporated advanced mixing techniques, CutMix and MixUp, which blend image patches and label vectors to improve model robustness and mitigate overfitting, especially important given the modest original dataset size.

Both models were fine‑tuned using ImageNet pre‑trained weights. Training hyper‑parameters were largely identical: batch size = 32, learning rate = 1e‑4, Adam optimizer, early stopping with patience = 10, and model checkpointing. InceptionV3 was trained for 100 epochs, while ResNet50 ran for 34 epochs. The authors also split the data into training, validation, and test subsets (approximately 70 %/20 %/10 %).

Performance metrics reveal a clear advantage for InceptionV3. It achieved a near‑perfect training accuracy (~100 %) and an outstanding validation accuracy of 99.96 %, culminating in an overall test accuracy of 98.15 % (372/379 correct predictions). The confusion matrix shows perfect classification for Anthracnose (170/170) and Fruit fly (118/118) and only 7 Healthy samples mis‑identified as Fruit fly. SHAP (Shapley Additive Explanations) visualizations highlighted that the model’s decisions were driven by disease‑specific visual cues such as discoloration and lesion patterns, providing interpretability that is crucial for agricultural stakeholders.

ResNet50, despite its deeper architecture and residual connections, lagged behind. Training accuracy stayed above 90 %, but validation accuracy fluctuated around 80 % and occasionally dropped below 40 %, indicating instability and overfitting. Test accuracy settled at 94.46 % (363/377). The confusion matrix shows modest misclassifications: 4 Anthracnose samples labeled Healthy, 12 Fruit fly samples mislabeled (10 as Anthracnose, 2 as Healthy), and 5 Healthy samples misidentified.

The authors discuss why InceptionV3 outperforms ResNet50 in this context. InceptionV3’s multi‑scale inception modules efficiently capture both fine‑grained texture and broader shape information, which is well‑suited to the subtle visual differences among guava disease symptoms. Its use of 1×1 convolutions reduces dimensionality, limiting parameter count and helping to avoid overfitting on a relatively small dataset. ResNet50’s depth, while advantageous for large‑scale image tasks, appears excessive for the limited data, leading to higher variance and poorer generalization.

The study also underscores the value of data mixing (CutMix, MixUp) for agricultural image classification, demonstrating measurable gains in both models. Moreover, the inclusion of SHAP (and mention of LIME) provides a transparent AI pipeline, allowing agronomists to verify that the model focuses on biologically relevant regions rather than spurious background artifacts.

In conclusion, the paper validates that a transfer‑learned InceptionV3, combined with modern augmentation and explainable AI tools, delivers high‑accuracy, reliable guava disease detection suitable for deployment on edge devices or mobile platforms. Future work suggested includes model quantization, further reduction of computational footprint, and extensive field testing under variable lighting and background conditions to ensure robustness in real‑world farming environments.