Fine-Tuning Strategies of EfficientNetV2B0 for Fish Freshness Classification Based on Eye Images with Grad-CAM
DOI:
10.59888/ajosh.v1i04.788Published:
2023-01-25Downloads
Abstract
Background: Using eye images to classify fish freshness requires an adaptation configuration of a pretrained model that suits the target data, yet the effect of fine-tuning depth on compact backbones remains rarely examined.
Objective: This study analysed four training configurations of EfficientNetV2B0 and interpreted their predictions using Grad-CAM.
Method: Version 2 of the Multistage Fish Eyes Dataset contains 4,800 images in three classes (Fresh, Highly Fresh, Not Fresh), partitioned into 3,360 training, 720 validation, and 720 test images. The configurations were a frozen backbone and the unfreezing of approximately the last 20%, the last 50%, and the whole backbone, each with ImageNet weights, 224 × 224-pixel inputs, and one seed.
Results: Test accuracy reached 82.36%, 96.67%, 97.22%, and 87.92%, respectively. Unfreezing approximately the last 50% of layers produced the lowest validation loss and the highest macro F1 (97.23%), with only four more correct predictions than the 20% configuration, a non-significant difference (exact McNemar p ? 0.125). Grad-CAM overlays showed attribution on the eye and surrounding head region but did not include the Not Fresh class.
Conclusion: Partial fine-tuning is a promising and computationally economical option on the partition used, although the findings remain exploratory because one seed was used, training factors changed together, and fish identity separation was not audited.
Keywords:
fish freshness EfficientNetV2B0 fine-tuning transfer learning Grad-CAMLicense
Copyright (c) 2023 Tio Bagus Purnama Apriliano
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.
