REVIEW 6 major objections 6 minor 24 references
Image Quality Enhancement and Detection of Small and Dense Objects in Industrial Recycling Processes
T0 review · 6 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read This paper claims that radiation-degraded industrial images can be restored by a lightweight two-stage network, DDSRNet, and that on a new 10,000-image industrial dataset the largest YOLOv8 and YOLO11 variants detect small, dense, overlappi
desk verdict New industrial dataset and detector benchmark are the real contribution; the DDSRNet noise model is mislabeled and the enhancement claims are untested. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing piece is DDSRNet, a shallow fully convolutional network with two stages—denoising first, super-resolution second—linked by skip connections and PixelUnshuffle/PixelShuffle down/upsampling, trained with a dynamic loss that spends the first epochs on denoising and gradually shifts weight to super-resolution. A secondary mechanism is the bounding-box diagonal as a proxy for object length, combined with the interquartile range to flag anomalies without camera calibration.
What would settle it
Take a camera feed from the actual radiation-exposed recycling line and a clean reference of the same scene; add the paper's synthetic noise to the clean reference, run DDSRNet on both the synthetic-noisy and real-noisy frames, and compare restored quality. If PSNR/SSIM gains on synthetic noise do not appear on real radiation noise, the central enhancement claim fails. Separately, evaluating the same detectors on the paper's 50,000-image synthetic dataset would test whether YOLOv8-x/YOLO11-x remain best at larger scale.
Extended reading notes
Core claim
The central claim is that in a radiation-heavy industrial recycling setting, a purpose-built dataset and two existing deep-learning tools are enough: a lightweight two-stage convolutional network called DDSRNet can denoise and 4x super-resolve degraded images, and the large YOLOv8 and YOLO11 variants, especially their -x versions, give the most reliable detection of small, dense, overlapping objects among the tested models. The paper also claims that a simple interquartile-range rule applied to bounding-box diagonals can flag long or oversized objects with up to 93% accuracy, avoiding expensive segmentation.
Load-bearing premise
The synthetic noise model (Poisson shot noise plus salt-and-pepper noise) adequately approximates how radiation actually degrades camera sensors; if it does not, DDSRNet's training and evaluation are disconnected from real deployment.
Editorial extensions
If this is right
- Deployment in a recycling plant can use YOLOv8-x or YOLO11-x at 640x640 as a default detector, with 1280x1280 variants available when small-object recall matters more than speed.
- DDSRNet can serve as a preprocessing stage that improves both human viewing and downstream detection without adding heavy compute.
- The IQR anomaly rule can flag oversized or elongated objects with up to 93% accuracy, independent of camera calibration, so it can be deployed across different camera setups.
- The bounding-box diagonal provides a fast length estimate, but error grows with object length, so precision-critical measurements still require segmentation.
Reading between the lines
- Because DDSRNet is tested only on synthetic Poisson and salt-and-pepper noise, its transfer to real radiation noise is the main open question; a direct field test would settle it.
- The 50,000-image, 6.5-million-instance synthetic dataset mentioned but not evaluated could become the more important asset: benchmarking detectors at that scale would reveal whether the YOLO ranking is stable or an artifact of the 10k-image test set.
- The IQR-on-diagonal idea generalizes beyond hulls: any dense-object stream with a stable length distribution could use the same outlier rule to flag anomalies without retraining.
- The dynamic-loss schedule suggests a wider recipe: prioritize the easier restoration task first, then shift to the harder task, which may transfer to other joint low-level vision problems.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper addresses two problems in an industrial nuclear-recycling setting: detecting small, dense, and overlapping objects, and enhancing noisy images. It introduces a new dataset (UDD) of 10k images with over 120k bounding-box annotations, compares a range of object detectors (YOLO variants, Faster R-CNN, RT-DETR), proposes DDSRNet, a lightweight joint denoising and 4x super-resolution model trained with a composite L1/SSIM/PSNR loss, and uses the interquartile range (IQR) over bounding-box diagonals to detect anomalous long hulls. The central claims are that DDSRNet effectively mitigates radiation-induced noise, that it improves downstream detection, and that the IQR method detects anomalies with up to 93% accuracy.
Significance. The UDD dataset, if released with proper splits and documentation, could be a useful benchmark for small-object detection in industrial environments. The detector comparison, despite missing details, provides a useful first reference point for practitioners. DDSRNet's two-stage architecture and dynamic loss weighting are reasonable ideas, but the paper does not currently establish their value because it lacks baseline comparisons, uses a nonstandard shot-noise model, and does not test the claimed benefit on downstream detection. The provision of code and dataset links is a positive reproducibility gesture, but the claims in the paper substantially exceed the evidence. If the identified gaps are addressed, the work could be a solid application-oriented contribution.
major comments (6)
- [Section 2, Eq. (1)] The shot-noise model in Eq. (1) is not standard Poisson shot noise: each pixel is replaced by a Poisson sample only with probability P and otherwise left unchanged. With P up to 0.2, at least 80% of pixels remain untouched, producing sparse impulse-like corruption rather than signal-dependent noise on every pixel. Thus the 'Shot noise' rows in Table 1 do not measure performance under actual shot noise, and the conclusion that DDSRNet 'effectively mitigates the effects of radiation-induced noise' (Section 2.3, Conclusion) is not supported by the reported experiments. Please either adopt the standard model I_noisy ~ Poisson(I) for all pixels, or justify the mixture model with physical sensor measurements, and ideally validate on real radiation-affected images.
- [Section 2.3, Table 1] DDSRNet is evaluated without any baseline comparisons. Reporting absolute PSNR/SSIM on Set5, Set14, BSD100, and UDD does not establish 'competitive performance' or 'generalization capability.' Please compare with at least one classical denoiser (e.g., BM3D) and one recent deep SR/denoising network under identical noise and scale settings, and specify the training/evaluation protocol for each dataset (e.g., which images are used for training DDSRNet before testing on standard SR benchmarks).
- [Section 2.3, last paragraph] The paragraph asserts that 'the proposed model enhances the performance of downstream object detection tasks.' No experiment in the paper feeds enhanced images to a detector and compares against the original images. This claim is load-bearing for the industrial motivation. Please add a quantitative experiment, e.g., reporting mAP on UDD with and without DDSRNet preprocessing, or remove the claim.
- [Section 3.3, Fig. 4] The anomaly detection results are summarized as 'achieving up to 93% accuracy' (Section 3.4) and 'up to 93% accuracy' in the conclusion, but the per-length detection accuracies in Fig. 4 range from 23.33% to 93.33% over only 30 repetitions, and the text states an overall detection rate of 72%. Reporting the best-case value in the conclusion is misleading. Please report detection accuracy per length with confidence intervals, an overall ROC curve or precision-recall summary, and the false-positive rate on the 2,000 normal hulls.
- [Section 3.2, Table 2] The object-detector comparison lacks essential training details: no epochs, batch size, optimizer, learning rate, pretrained initialization, or augmentation protocol. Furthermore, several entries appear inconsistent with the metric definitions: e.g., YOLOv5x has mAP@0.5 = 42.6 while mAP0.5-0.95 = 50.7, and YOLOv7x has mAP@0.5 = 38.0 with mAP0.5-0.95 = 53.1; mAP@0.5 should generally be at least as large as mAP0.5-0.95. The 'COCO-17' column is unexplained. Please correct the table and provide the experimental protocol.
- [Section 2.2, Eq. (3)] The training loss in Eq. (3) includes a PSNR term (100/PSNR) and the evaluation in Section 2.3 uses PSNR and SSIM. This is not necessarily fatal because the benchmarks provide external grounding, but the reported numbers partly reflect the training objective. Please discuss this overlap and, ideally, include an independent evaluation (e.g., on noise levels not seen in training, or with additional perceptual/edge-preservation metrics).
minor comments (6)
- [Abstract and Section 2] The abstract says 'fully connected convolutional network'; this is likely intended as 'fully convolutional network.' Please clarify.
- [Section 3.2, text] The model list reads 'YOLOv9, YOLO-NAS, YOLOv9 and YOLO11' — YOLOv9 appears twice; the second should likely be YOLOv10.
- [Section 3.3, Fig. 4] The text refers to 'Table 4' but no Table 4 exists; the data appears as a table inside Fig. 4. Please renumber or reference correctly.
- [Section 2.2] The training paragraph says the model was 'trained and tested separately for each noise type,' but it does not specify whether the same UDD images are used for training and testing, or whether the standard benchmarks are used only for evaluation. Please state the exact split and protocol.
- [Section 3.1] The dataset description would benefit from a class list, object-size statistics, annotation protocol, and a defined train/validation/test split. This is important for reproducibility and for interpreting the detector comparison.
- [Conclusion] The conclusion calls DDSRNet a 'shallow model' while Section 2.1 describes multiple convolutional blocks and skip connections; please quantify complexity (parameters, FLOPs, latency) to justify 'shallow' and 'real-time.'
Circularity Check
No significant circularity; minor overlap between DDSRNet's training loss and evaluation metrics does not reduce to a fit.
full rationale
The paper's main derivation chain is self-contained against external benchmarks. DDSRNet is trained on Set5/Set14/BSD100 and a held-out 20% split of UDD, with the quantitative evaluation using PSNR/SSIM on those test images. Although Eq. (3) includes PSNR and SSIM terms in the training loss and Table 1 reports PSNR/SSIM, this is an objective/metric overlap, not a circular prediction: the reported numbers are on separate test inputs, and the benchmarks provide external grounding. The IQR anomaly detector in Sec. 3.4 is an unsupervised rule applied to the same stack it characterizes; its 'detection accuracy' is in-sample descriptive rather than a held-out prediction, but it is not a fitted parameter renamed as a prediction, and it is not the paper's central claim. The noise model in Eq. (1) is a sparse mixture rather than standard Poisson shot noise and is used for both training and evaluation, which raises a generalization/correctness concern about the claim of mitigating real radiation-induced noise, but this is not a circularity: the model is not defined in terms of the evaluation result, and no self-citation or imported-uniqueness argument is load-bearing. Accordingly, no formal circular step is identified.
Assumptions & free parameters
free parameters (3)
- noise probabilities (P, ps, pp) =
P in [0,0.2], ps/pp in [0,0.1]
- loss weighting schedule (lambda, beta) =
lambda from 1 to 0.5, beta from 0 to 0.5 in steps of 0.1 every 10 epochs
- architecture repeat counts R1 and R2 =
R1=2, R2=2
assumptions (3)
- domain assumption Poisson shot noise and salt-and-pepper noise approximate radiation-induced sensor degradation
- domain assumption The simulated table replicates real-world sorting conditions closely enough for detector transfer
- domain assumption Bounding box diagonal is roughly proportional to hull length
Cite this review
Pith. "Pith review of Image Quality Enhancement and Detection of Small and Dense Objects in Industrial Recycling Processes." pith.science (2026). https://pith.science/paper/EMQABEBB
@misc{pith2026250901332,
author = {Pith},
title = {Pith review of: Image Quality Enhancement and Detection of Small and Dense Objects in Industrial Recycling Processes},
year = {2026},
howpublished = {\url{https://pith.science/paper/EMQABEBB}},
note = {Machine review of arXiv:2509.01332}
}
read the original abstract
This paper tackles two key challenges: detecting small, dense, and overlapping objects (a major hurdle in computer vision) and improving the quality of noisy images, especially those encountered in industrial environments. [1, 2]. Our focus is on evaluating methods built on supervised deep learning. We perform an analysis of these methods, using a newly developed dataset comprising over 10k images and 120k instances. By evaluating their performance, accuracy, and computational efficiency, we identify the most reliable detection systems and highlight the specific challenges they address in industrial applications. This paper also examines the use of deep learning models to improve image quality in noisy industrial environments. We introduce a lightweight model based on a fully connected convolutional network. Additionally, we suggest potential future directions for further enhancing the effectiveness of the model. The repository of the dataset and proposed model can be found at: https://github.com/o-messai/SDOOD, https://github.com/o-messai/DDSRNet
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION Object detection algorithms are a critical technology in many industrial applications, where they enable automation, qual- ity control, and improved decision-making processes [1]. Al- though significant advances have been made in object de- tection algorithms through the application of deep learning techniques, the detection of small and over...
work page Pith review arXiv 2025
-
[2]
This damage can result in dead pixels that consistently output a fixed value
IMAGE QUALITY ENHANCEMENT Uranium radiation primarily induces localized energy depo- sition in camera sensors, leading to increased pixel values or, in severe cases, permanent pixel damage. This damage can result in dead pixels that consistently output a fixed value. Simulating the effects of uranium radiation on camera sensors requires advanced noise mod...
-
[3]
OBJECT DETECTION 3.1. New dataset The dataset for this study was collected using a simula- tion model developed in accordance with recommendations from Orano group. The simulated environments replicate real-world conditions as closely as possible. To account for variations in lighting, two types of light projection were used during image capture. The data...
work page 2000
-
[4]
The results from SOTA object detection methods are promising
CONCLUSION AND PERSPECTIVES A new dataset has been created to address the challenge of de- tecting small, dense, and overlapping objects in industrial en- vironments. The results from SOTA object detection methods are promising. The YOLOv8 and YOLOv11 series, particu- larly the -x variants, demonstrate the best trade-off between recall, precision, and par...
-
[5]
Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,
Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) , June 2023, pp. 7464–7475
work page 2023
-
[6]
Object detection with deep learning: A re- view,
Zhong-Qiu Zhao, Peng Zheng, Shou-tao Xu, and Xin- dong Wu, “Object detection with deep learning: A re- view,” IEEE transactions on neural networks and learn- ing systems, vol. 30, no. 11, pp. 3212–3232, 2019
work page 2019
-
[7]
Object detection in 20 years: A sur- vey,
Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye, “Object detection in 20 years: A sur- vey,” Proceedings of the IEEE, vol. 111, no. 3, pp. 257– 276, 2023
work page 2023
-
[8]
Jakaria Rabbi, Nilanjan Ray, Matthias Schubert, Subir Chowdhury, and Dennis Chao, “Small-object detec- tion in remote sensing images with end-to-end edge- enhanced gan and object detector network,” Remote Sensing, vol. 12, no. 9, pp. 1432, 2020
work page 2020
Show all 24 references
-
[9]
Faster r-cnn: Towards real-time object detection with region proposal networks,
Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” Advances in neural in- formation processing systems, vol. 28, 2015
2015
-
[10]
On single image scale-up using sparse-representations,
Roman Zeyde, Michael Elad, and Matan Protter, “On single image scale-up using sparse-representations,” in Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. Springer, 2012, pp. 711–730
2010
-
[11]
Finally, we select 100 images (80% for train, and 20% for test) from our 10k UDD dataset for evaluation
contains 100 images, primarily depicting natural land- scapes, food, and people. Finally, we select 100 images (80% for train, and 20% for test) from our 10k UDD dataset for evaluation. For noise simulations, we added two types of noise—Salt & Pepper (S&P) and Shot noise—to al...
-
[12]
Faster r-cnn: Towards real-time object detection with region proposal networks,
Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 6, pp. 1137–1149, 2016
2016
-
[13]
Image denoising: The deep learning revolution and be- yond—a survey paper,
Michael Elad, Bahjat Kawar, and Gregory Vaksman, “Image denoising: The deep learning revolution and be- yond—a survey paper,” SIAM Journal on Imaging Sci- ences, vol. 16, no. 3, pp. 1594–1654, 2023
2023
-
[14]
Photon, poisson noise,
Samuel W Hasinoff, “Photon, poisson noise,” in Com- puter vision: a reference guide , pp. 980–982. Springer, 2021
2021
-
[15]
Salt-and-pepper noise removal by median-type noise detectors and detail-preserving regularization,
Raymond H Chan, Chung-Wa Ho, and Mila Nikolova, “Salt-and-pepper noise removal by median-type noise detectors and detail-preserving regularization,” IEEE Transactions on image processing , vol. 14, no. 10, pp. 1479–1485, 2005
2005
-
[16]
Contour detection and hierarchical im- age segmentation,
Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik, “Contour detection and hierarchical im- age segmentation,” IEEE transactions on pattern anal- ysis and machine intelligence , vol. 33, no. 5, pp. 898– 916, 2010
2010
-
[17]
ultralyt- ics/yolov5: v7. 0-yolov5 sota realtime instance segmen- tation,
Glenn Jocher, Ayush Chaurasia, Alex Stoken, Jirka Borovec, Yonghye Kwon, Kalen Michael, Jiacong Fang, Zeng Yifu, Colin Wong, Diego Montes, et al., “ultralyt- ics/yolov5: v7. 0-yolov5 sota realtime instance segmen- tation,” Zenodo, 2022
2022
-
[18]
Yolov6: A single-stage object detection framework for industrial applications,
Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al., “Yolov6: A single-stage object detection framework for industrial applications,” arXiv preprint arXiv:2209.02976, 2022
2022 arXiv
-
[19]
A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas,
Juan Terven, Diana-Margarita C ´ordova-Esparza, and Julio-Alejandro Romero-Gonz ´alez, “A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas,” Machine Learning and Knowledge Extraction , vol. 5, no. 4, pp. 1680–1716, 2023
2023
-
[20]
Yolov9: Learning what you want to learn us- ing programmable gradient information,
Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao, “Yolov9: Learning what you want to learn us- ing programmable gradient information,”arXiv preprint arXiv:2402.13616, 2024
2024 arXiv
-
[21]
Yolov10: Real-time end-to-end object detection,
Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding, “Yolov10: Real-time end-to-end object detection,” arXiv preprint arXiv:2405.14458, 2024
2024 arXiv
-
[22]
Ultralytics yolo11,
Glenn Jocher and Jing Qiu, “Ultralytics yolo11,” https://github.com/ultralytics/ultralytics, 2024
2024
-
[23]
Detrs beat yolos on real-time object detection,
Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen, “Detrs beat yolos on real-time object detection,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2024, pp. 16965– 16974
2024
-
[24]
Segment anything,
Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al., “Segment anything,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026
2023
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.