Comparison of EfficientNetV2-S and ConvNeXt V2-Tiny for Potato Leaf Condition Classification with Grad-CAM
DOI:
10.59888/ajosh.v1i01.786Published:
2022-01-25Downloads
Abstract
Background: Variations in background, lighting, and viewing angle make potato leaf condition classification difficult in uncontrolled environments.
Objective: This study compared EfficientNetV2-S and ConvNeXt V2-Tiny on seven classes in terms of classification performance and inference efficiency and examined the image regions contributing to predictions using Gradient-weighted Class Activation Mapping (Grad-CAM).
Method: Of 3,076 initial images, 12 with label conflicts were quarantined and 49 identical copies removed, leaving 3,015 images, split into 2,412 training, 301 validation, and 302 test images with similar images kept in the same partition. Both models were trained using transfer learning, weighted cross-entropy, and three seeds.
Results: EfficientNetV2-S achieved a mean test accuracy of 83.22% and macro-F1 of 79.86%, whereas ConvNeXt V2-Tiny achieved 87.97% and 85.21%. The macro-F1 difference of 5.34 percentage points had a 95% confidence interval of 2.43–8.50 percentage points based on a paired group bootstrap, although the per-run advantage ranged from 1.06 to 11.06 points and the seed-42 McNemar test was not significant (p ? 0.39). On an NVIDIA T4 GPU, ConvNeXt V2-Tiny required 11.70 versus 25.39 ms per image. Grad-CAM visualizations showed different activation spreads for the two models on example test images.
Conclusion: ConvNeXt V2-Tiny delivered higher performance and lower latency in this internal test and is the more practical candidate for server-based diagnosis, although cross-location evaluation and spatial validation against lesion annotations are still required.
Keywords:
image classification potato leaf EfficientNetV2-S ConvNeXt V2-Tiny Grad-CAMLicense
Copyright (c) 2022 Yoga Pamungkas
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Authors who publish with this journal agree to the following terms:
- Authors retain copyright and grant the journal right of first publication with the work simultaneously licensed under a Creative Commons Attribution-ShareAlike 4.0 International. that allows others to share the work with an acknowledgement of the work's authorship and initial publication in this journal.
- Authors are able to enter into separate, additional contractual arrangements for the non-exclusive distribution of the journal's published version of the work (e.g., post it to an institutional repository or publish it in a book), with an acknowledgement of its initial publication in this journal.
- Authors are permitted and encouraged to post their work online (e.g., in institutional repositories or on their website) prior to and during the submission process, as it can lead to productive exchanges, as well as earlier and greater citation of published work.
