Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors utilized the B-Res41-ASE evaluation to classify images of silkworm cocoons and achieved desirable results effectively addressing some of the practical challenges in the cocoon industry. This provides an important reference for classifying cocoons with machines instead of manual labor and the study is innovative. However, modifications are required for publication in agricultural journals.

1. Extended Introduction Literature review, especially research relevant to this study

2.  2.2.3. the interpretation of mathematical formula equations in the feature fusion module could be more intuitive

3.  Figure 10: Training curves for different feature fusion enhancements Can you zoom in locally, as the colors are similar, so the visual representation is not very clear?

4. For the section on comparative testing, is the model for comparison too old? Have the authors tried the latest models such as MobileNet and ShuffleNet latest models?

5. The discussion section needs to highlight the progress by emphasizing the comparison between the author's approach and existing approaches.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper introduces ResNet41 as the backbone and integrates bilinear pooling, ASFF, and SE modules, but it does not sufficiently justify why these specific components were chosen over others. For instance, why ResNet41 instead of other advanced architectures like EfficientNet or Vision Transformers (ViTs)? A more detailed comparison or rationale is necessary. The paper emphasizes performance improvements on the given dataset but does not evaluate the model's robustness across diverse datasets or unseen conditions. This is particularly critical for practical deployment, where lighting, background, or cocoon types may vary.  While the proposed model outperforms others in terms of accuracy, the paper does not provide insights into its computational efficiency. For practical implementation in sorting systems, metrics like inference time, memory usage, and hardware requirements are essential but not discussed. The dataset is augmented to balance class distribution, but there is no explicit evaluation of potential bias in the original dataset. Does the model perform equally well across all classes in non-augmented scenarios? This could influence real-world reliability. While comparisons with AlexNet, GoogLeNet, and other older architectures are made, the exclusion of more recent fine-grained classification models (e.g., transformer-based models or YOLOv8) weakens the claim of achieving state-of-the-art performance. The paper does not thoroughly analyze the challenges posed by fine-grained image classification, such as intra-class variability and inter-class similarity. Providing examples of misclassifications and their root causes would offer more transparency. While the paper presents a promising approach to silkworm cocoon classification with competitive performance, its contributions are undermined by the lack of rigorous justification for model choices, limited evaluation of real-world applicability, and insufficient comparison with recent state-of-the-art methods. Figures like 3, 6, and 14 are dense and challenging to interpret at a glance. Consider simplifying the visualizations or providing detailed legends to aid understanding.  The paper is well-written and addresses a significant problem in a novel way. Addressing the above suggestions will strongly contribute to fine-grained image classification and the automation of silkworm cocoon sorting.

Comments on the Quality of English Language

Improvement required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf