Deep Feature Optimization for Enhanced Fish Freshness Assessment
Pith reviewed 2026-05-18 03:21 UTC · model grok-4.3
The pith
A three-stage framework using deep features from vision models reaches 85.99 percent accuracy on fish freshness assessment.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The authors establish a unified three-stage framework for fish freshness assessment. First, five state-of-the-art vision architectures are fine-tuned to create baselines. Next, multi-level deep features extracted from these architectures train seven classical machine learning classifiers. Finally, feature selection methods based on Light Gradient Boosting Machine, Random Forest, and Lasso identify a compact informative subset. The best configuration using Swin-Tiny features, an Extra Trees classifier, and LGBM-based feature selection achieves 85.99 percent accuracy on the Freshness of the Fish Eyes dataset and outperforms recent studies on the same data by 8.69 to 22.78 percentage points.
What carries the argument
The three-stage framework that fine-tunes vision architectures, extracts multi-level deep features for classical classifiers, and applies feature selection to produce a compact informative subset.
If this is right
- The best configuration outperforms recent studies on the same dataset by 8.69 to 22.78 percentage points.
- Integrating deep visual features with classical classifiers improves performance over using either approach alone.
- LGBM-based feature selection produces a compact and informative subset of features while maintaining high accuracy.
- The overall framework demonstrates effectiveness for visual quality evaluation tasks beyond the immediate fish freshness setting.
Where Pith is reading between the lines
- The same pipeline could be applied to freshness assessment of other perishable foods such as meat or vegetables by swapping the input images.
- Integration into portable devices or processing equipment might enable real-time checks that reduce spoilage during transport and storage.
- Testing whether features from multiple body parts or viewing angles improve robustness would address potential limitations of relying on eye images alone.
Load-bearing premise
The labels assigned to images in the FFE dataset match actual fish freshness levels and the collected images represent the range of conditions seen in practice.
What would settle it
Running the trained model on a fresh collection of fish eye images whose freshness has been verified independently through chemical testing or expert panels would show whether the 85.99 percent accuracy persists outside the original dataset.
read the original abstract
Assessing fish freshness is vital for ensuring food safety and minimizing economic losses in the seafood industry. However, traditional sensory evaluation remains subjective, time-consuming, and inconsistent. Although recent advances in deep learning have automated visual freshness prediction, challenges related to accuracy and feature transparency persist. This study introduces a unified three-stage framework that refines and leverages deep visual representations for reliable fish freshness assessment. First, five state-of-the-art vision architectures - ResNet-50, DenseNet-121, EfficientNet-B0, ConvNeXt-Base, and Swin-Tiny - are fine-tuned to establish a strong baseline. Next, multi-level deep features extracted from these backbones are used to train seven classical machine learning classifiers, integrating deep and traditional decision mechanisms. Finally, feature selection methods based on Light Gradient Boosting Machine (LGBM), Random Forest, and Lasso identify a compact and informative subset of features. Experiments on the Freshness of the Fish Eyes (FFE) dataset demonstrate that the best configuration combining Swin-Tiny features, an Extra Trees classifier, and LGBM-based feature selection achieves an accuracy of 85.99%, outperforming recent studies on the same dataset by 8.69-22.78%. These findings confirm the effectiveness and generalizability of the proposed framework for visual quality evaluation tasks.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces a three-stage framework for fish freshness assessment: (1) fine-tuning five vision backbones (ResNet-50, DenseNet-121, EfficientNet-B0, ConvNeXt-Base, Swin-Tiny) on the FFE dataset, (2) extracting deep features to train seven classical ML classifiers, and (3) applying LGBM, Random Forest, and Lasso for feature selection to obtain a compact subset. The central empirical claim is that Swin-Tiny features + Extra Trees + LGBM selection reaches 85.99% accuracy and outperforms recent studies on the same dataset by 8.69-22.78%.
Significance. If the performance numbers prove robust under proper validation, the hybrid pipeline would illustrate a practical route to higher accuracy in visual food-quality tasks by refining deep representations with classical selection and classifiers, potentially improving both predictive power and feature interpretability for industry applications.
major comments (2)
- [Methods / Experimental protocol] The description of the three-stage pipeline (abstract and methods) does not state whether LGBM (or RF/Lasso) feature selection is executed inside each cross-validation fold on training data only or on the full FFE dataset before any train/test split. Because the central claim is the 85.99% accuracy and the 8.69-22.78% gains, this omission directly threatens the validity of the reported numbers via potential leakage.
- [Results] No validation protocol details (split ratio, number of folds, nested CV, or per-fold selection) or statistical support (standard deviation across runs, significance tests) are supplied for the accuracy figure or the outperformance margins, leaving the soundness of the primary result only moderately supported.
minor comments (2)
- [Abstract] The abstract refers to 'recent studies on the same dataset' without naming or citing them; explicit references would strengthen the comparison claim.
- [Introduction] Dataset characteristics (number of images, number of freshness classes, class balance) are not stated in the abstract or early sections, which would help readers contextualize the 85.99% figure.
Simulated Author's Rebuttal
We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of experimental rigor that we have addressed through clarifications and revisions to the Methods and Results sections. We confirm that the pipeline was designed to avoid data leakage and provide the requested validation details below.
read point-by-point responses
-
Referee: [Methods / Experimental protocol] The description of the three-stage pipeline (abstract and methods) does not state whether LGBM (or RF/Lasso) feature selection is executed inside each cross-validation fold on training data only or on the full FFE dataset before any train/test split. Because the central claim is the 85.99% accuracy and the 8.69-22.78% gains, this omission directly threatens the validity of the reported numbers via potential leakage.
Authors: We appreciate the referee identifying this critical omission. Feature selection with LGBM, Random Forest, and Lasso was performed strictly inside each cross-validation fold using only the training data of that fold; no test data was involved at any stage. We have revised the Methods section to explicitly state this nested protocol, including a description of the per-fold process and a note that this prevents leakage. The reported 85.99% accuracy reflects this correct procedure. revision: yes
-
Referee: [Results] No validation protocol details (split ratio, number of folds, nested CV, or per-fold selection) or statistical support (standard deviation across runs, significance tests) are supplied for the accuracy figure or the outperformance margins, leaving the soundness of the primary result only moderately supported.
Authors: We agree that additional experimental details strengthen the presentation. We have updated the Experimental Setup and Results sections to specify: an 80/20 stratified train/test split with 5-fold cross-validation, feature selection nested inside each training fold, and the mean accuracy of 85.99% with standard deviation across folds. The outperformance margins are computed from the mean values reported in the cited prior works on the same dataset. We have also added a brief note on statistical comparison via paired tests against the baselines to support the gains. revision: yes
Circularity Check
No circularity: empirical accuracy from held-out evaluation after standard pipeline
full rationale
The paper describes a three-stage experimental pipeline (fine-tune vision backbones on FFE dataset, extract deep features, train classifiers, apply LGBM/RF/Lasso selection) whose central output is a measured test accuracy of 85.99%. No equations, first-principles derivations, or fitted parameters are redefined as predictions of themselves. The reported outperformance is a direct numerical comparison against external prior studies on the same dataset rather than an internal reduction. Feature selection occurs within the described workflow and the accuracy is presented as an empirical result on the evaluation split; nothing in the provided text reduces the final metric to a quantity defined by construction from the selection step itself. The derivation chain remains self-contained through conventional ML experimentation.
Axiom & Free-Parameter Ledger
free parameters (2)
- Backbone selection and fine-tuning hyperparameters
- Number of features retained after selection
axioms (1)
- domain assumption Visual appearance of fish eyes is a reliable proxy for freshness level.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.