Explainable CNN benchmark for snake fruit ripeness classification using grad-CAM visualization
Abstract
Snake fruit ripeness classification lacks both a standardized image dataset and a systematic multi-architecture convolutional neural network (CNNs) benchmark. This study addresses both gaps by constructing a dataset of 741 images (ripe and unripe, class-balanced in training and validation, imbalanced in test by design) through web scraping and large language model (LLM)-assisted annotation with confidence-based quality filtering, and by benchmarking five transfer learning architectures (VGG16, MobileNetV2, EfficientNetB0, DenseNet121, and ResNet50) with gradient-weighted class activation mapping (Grad-CAM) explainability on metrics relevant to field deployment, reporting results as mean ±95% confidence interval over five random seeds. DenseNet121 achieved the highest mean test accuracy (91.29%±1.67, F1: 0.9135±0.0166), closely followed by MobileNetV2 (90.82%±2.61) with the smallest model size (8.7 MB) and fastest practical inference among the mid-accuracy architectures (0.18 ms); confidence intervals overlap substantially across all five architectures, indicating no single architecture is statistically clearly superior on accuracy alone. Grad-CAM maps confirm that models attend to fruit surface texture rather than background. Dataset, trained models, and the annotation and evaluation pipeline are publicly released to support reproducible research on snake fruit quality assessment.
Keywords
Convolutional neural network; Food security; Gradient-weighted class activation mapping; Ripeness classification; Snake fruit
Full Text:
PDFDOI: https://doi.org/10.11591/eei.v15i6.13354
Refbacks
- There are currently no refbacks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Bulletin of Electrical Engineering and Informatics (BEEI)
ISSN: 2089-3191
,
e-ISSN: 2302-9285
This journal is published by the
Institute of Advanced Engineering and Science (IAES)
.