具有並列連接的最新卷積神經網絡，例如 DenseNet，FractalNet 和
HarDNet 藉由大量的特徵融合運算在眾多計算機視覺任務（包括圖像分類
和目標檢測）取得了傲人的成績。但是，相對密集的連接模式限制了這些
有效模型的推理速度。在本文中，我們提出了一個維持多尺度特徵融合特
色且但具有更少的記憶體搬運量的新計算模塊，稱為 HarDX 塊和一個用於
設計特定於輸入大小的卷積神經網路結構名為 Stage Receptive Field 的新穎
概念。基於這兩個創新的概念，我們提出了一種新的卷積神經網路結構，
名為 HarDNeXt，速度分別比 DenseNets 和 HarDNets 快了 200％和 30％並
且在進行推論所需的功耗上分別少了 60％和 15％。對於結腸鏡檢查圖像息
肉分割的應用任務，HarDNeXt 與最先進的神經網路模型相比達到相同水平
的精度並將運行速度提高了 25％。

State­of­the­art Convolution Neural Networks with extensive feature aggregation via shortcut connections such as DenseNet, FractalNet and HarDNet have
achieved remarkable results with extensive feature aggregations in numerous computer vision tasks, including image classification and object detection . However, the relatively dense connection pattern has impaired the inference speed. We
propose a new computation block, called HarDX block, for maintaining multireceptive field feature fusion with low DRAM traffic and a novel concept, called
Stage Receptive Field, for designing input­size­specific network architectures. Accordingly, we propose a new CNN architecture named HarDNeXt, which is 200%
and 30% faster than DenseNets and HarDNets, respectively, while consuming 60%
and 15% less energy, respectively. For an application task of polyp segmentation
of colonoscopy images, HarDNeXt achieves the same level of accuracy and runs
25% faster compared with a state­of­the­art network. Code, data, and experiment
setup are open­sourced in GitHub.

Acknowledgements
摘要 i
Abstract ii
1 Introduction 1
2 Related Work 7
2.1 Manual Designed Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Neural Architecture Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Proposed Methods 11
3.1 Improving Harmonic Dense Connection . . . . . . . . . . . . . . . . . . . . . 11
3.2 Stage Receptive Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 HarDNeXt Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Experiment Results 21
4.1 Classification Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.1 Training Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Polyp Segmentation Application . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2.1 Dataset and Training Strategy . . . . . . . . . . . . . . . . . . . . . . 25
4.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5 Conclusion and Future Work 29
Bibliography 31

Bibliography
[1] Baker, B., Gupta, O., Naik, N., and Raskar, R. Designing neural network architectures
using reinforcement learning. CoRR abs/1611.02167 (2016).
[2] Bernal, J., Sánchez, F., Fernández­Esparrach, G., Gil, D., Rodríguez, C., and Vilariño, F.
Wm­dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency
maps from physicians. Computerized Medical Imaging and Graphics 43 (Jan. 2015), 99–
111.
[3] Cai, H., Zhu, L., and Han, S. Proxylessnas: Direct neural architecture search on target
task and hardware. CoRR abs/1812.00332 (2018).
[4] Chao, P., Kao, C., Ruan, Y., Huang, C., and Lin, Y. Hardnet: A low memory traffic
network. CoRR abs/1909.00948 (2019).
[5] Deng, J., Dong, W., Socher, R., Li, L., Kai Li, and Li Fei­Fei. Imagenet: A large­scale
hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern
Recognition (2009), pp. 248–255.
[6] Fan, D.­P., Ji, G.­P., Zhou, T., Chen, G., Fu, H., Shen, J., and Shao, L. Pranet: Parallel
reverse attention network for polyp segmentation, 2020.
[7] He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition.
CoRR abs/1512.03385 (2015).
[8] Howard, A., Sandler, M., Chu, G., Chen, L., Chen, B., Tan, M., Wang, W., Zhu, Y.,
Pang, R., Vasudevan, V., Le, Q. V., and Adam, H. Searching for mobilenetv3. CoRR
abs/1905.02244 (2019).
[9] Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto,
M., and Adam, H. Mobilenets: Efficient convolutional neural networks for mobile vision
applications. CoRR abs/1704.04861 (2017).
[10] Huang, C.­H., Wu, H.­Y., and Lin, Y.­L. Hardnet­mseg: A simple encoder­decoder polyp
segmentation neural network that achieves over 0.9 mean dice and 86 fps, 2021.
[11] Huang, G., Liu, Z., and Weinberger, K. Q. Densely connected convolutional networks.
CoRR abs/1608.06993 (2016).
[12] Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., de Lange, T., Johansen, D., and
Johansen, H. D. Kvasir­seg: A segmented polyp dataset, 2019.
31[13] Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25:
26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of
a meeting held December 3­6, 2012, Lake Tahoe, Nevada, United States (2012), pp. 1106–
1114.
[14] Larsson, G., Maire, M., and Shakhnarovich, G. Fractalnet: Ultra­deep neural networks
without residuals. CoRR abs/1605.07648 (2016).
[15] Lee, Y., Hwang, J., Lee, S., Bae, Y., and Park, J. An energy and gpu­computation efficient
backbone network for real­time object detection. CoRR abs/1904.09730 (2019).
[16] Liu, H., Simonyan, K., and Yang, Y. DARTS: differentiable architecture search. CoRR
abs/1806.09055 (2018).
[17] Sandler, M., Howard, A. G., Zhu, M., Zhmoginov, A., and Chen, L. Inverted residuals
and linear bottlenecks: Mobile networks for classification, detection and segmentation.
CoRR abs/1801.04381 (2018).
[18] Silva, J., Histace, A., Romain, O., Dray, X., and Granado, B. Toward embedded detection
of polyps in wce images for early diagnosis of colorectal cancer. International journal of
computer assisted radiology and surgery 9, 2 (March 2014), 283—293.
[19] Simonyan, K., and Zisserman, A. Very deep convolutional networks for large­scale image
recognition. arXiv preprint arXiv:1409.1556 (2014).
[20] Tajbakhsh, N., Gurudu, S., and Liang, J. Automated polyp detection in colonoscopy videos
using shape and context information. IEEE Transactions on Medical Imaging 35, 2 (Feb.
2016), 630–644. Publisher Copyright: © 2015 IEEE. Copyright: Copyright 2017 Elsevier
B.V., All rights reserved.
[21] Tan, M., and Le, Q. V. Efficientnet: Rethinking model scaling for convolutional neural
networks. CoRR abs/1905.11946 (2019).
[22] Vázquez, D., Bernal, J., Sánchez, F. J., Fernández­Esparrach, G., López, A. M., Romero,
A., Drozdzal, M., and Courville, A. C. A benchmark for endoluminal scene segmentation
of colonoscopy images. CoRR abs/1612.00799 (2016).
[23] Wu, B., Dai, X., Zhang, P., Wang, Y., Sun, F., Wu, Y., Tian, Y., Vajda, P., Jia, Y., and
Keutzer, K. Fbnet: Hardware­aware efficient convnet design via differentiable neural
architecture search. CoRR abs/1812.03443 (2018).
[24] Zhong, Z., Yan, J., and Liu, C. Practical network blocks design with q­learning. CoRR
abs/1708.05552 (2017).