HieraEdgeNet: A Multi-Scale Edge-Enhanced Framework for Automated Pollen Recognition

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents HieraEdgeNet, a multi-scale edge-enhanced detection framework tailored for automated pollen recognition. The work is well motivated, technically solid, and clearly demonstrates improvements over current state-of-the-art detectors. The experimental results are comprehensive and convincingly support the proposed approach. Overall, this is a strong contribution with practical relevance for palynology, biodiversity monitoring, and agricultural applications. Only minor revisions are needed before publication.

1. Although the ablation table is comprehensive, the manuscript would benefit from more detailed reasoning explaining why individual modules (HEM, SEF, CSPOKM) underperform when used alone. For example, describe how the absence of SEF causes edge priors to be underutilized, or why CSPOKM alone has limited impact without fused edge–semantic cues.
2. The manuscript does not clearly specify the exact learning rate schedule, optimizer parameters, batch size, and augmentation ratios. These parameters influence performance and should be briefly listed for reproducibility.
3. The authors freeze Sobel kernels, but do not explore or discuss whether learnable gradient kernels might yield stronger edge representations. A short justification or citation would strengthen the rationale.
4. HEM generates EP3, EP4, and EP5 at resolutions aligned to P3, P4, and P5, yet P5 in the main backbone is at 1/32 resolution. The manuscript should clarify how the spatial mismatch between EP5 (1/16) and P5 (1/32) is resolved or mitigated during fusion.
5. CSPOKM introduces large-kernel and multi-branch operations. Provide explaining why this computational overhead remains justified relative to observed accuracy gains in small-object contexts.
6. Focal Loss is used, but the authors do not state whether other strategies (e.g., class-balanced loss, re-weighting, or sample re-sampling) were considered. This omission should be briefly addressed, even if only to state that preliminary tests showed no improvement.
7. The pruning ratio and criteria (e.g., sensitivity-based, per-layer scoring) are not fully described. Add a brief statement of how pruning thresholds were determined to achieve the reported compression–accuracy balance.
8. The manuscript states that bounding box regression uses DFL “typically used with CIoU/GIoU/SIoU,” but does not specify which variant was used in experiments. Please identify the exact IoU loss used and whether alternatives were tested.
9. The paper does not specify the NMS threshold, confidence threshold (other than from F1 curve), or whether soft-NMS was explored. Since pollen objects can be close or overlapping, specifying this would improve clarity.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents HieraEdgeNet, a novel and well-engineered framework for automated pollen recognition. The authors address the significant challenges of detecting small, indistinct targets in complex microscopic backgrounds by proposing a synergistic architecture with three core modules: the Hierarchical Edge Module (HEM), the Synergistic Edge Fusion (SEF) module, and the Cross Stage Partial Omni-Kernel Module (CSPOKM). The research is motivated by a clear and important problem in palynology and related fields. The proposed architecture is innovative, the experimental validation is comprehensive and rigorous, and the results are highly convincing. The paper is well-written and logically structured.

While the manuscript is of high quality, I have the following suggestions that I believe will enhance its impact and rigor.

1. On the Use of a Fixed Sobel Operator in the HEM Module: The HEM module employs a fixed-weight Sobel convolution (SobelConv) for edge extraction. This is a strong inductive bias. While effective, it raises questions about the model’s ability to learn edge patterns that may not be perfectly captured by classical operators, potentially limiting its generalization to novel pollen morphologies. Suggestion: The authors should add a brief discussion in Section 2.1 to justify this design choice. Is it for computational efficiency, stability, or the assumption that pollen edges are well-represented by Sobel filters? At a minimum, this should be acknowledged as a potential limitation. An even stronger addition would be a small ablation experiment comparing the fixed SobelConv with a simple, learnable 3x3 convolution to quantify the trade-off.

2. Deeper Analysis of Computational Complexity:The paper rightly notes that the CSPOKM module introduces substantial computational cost and reports the overall GFLOPs. However, a breakdown of this cost is missing. Suggestion: In Section 3.2, it would be highly valuable to provide a brief complexity decomposition. For instance, the authors could estimate the GFLOPs contribution of each of the three core modules (HEM, SEF, CSPOKM) relative to the baseline. This would allow readers to better understand the “price-performance” trade-off of each innovation and assess which component offers the best efficiency gain.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

see the attachment.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf