[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-513532-105":59,"doc-detail-513532-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","puca-patch-unshuffle-and-channel-attention-for-enhanced-self-supervised-image-denoising-implementation-details","PUCA: Patch-Unshuffle and Channel Attention for Enhanced Self-Supervised Image Denoising - Implementation Details","","Supplementary material for the NeurIPS 2023 work PUCA presents training, network, and visualization implementation details for enhanced self-supervised image denoising. The method trains on SIDD-Medium and DND with 256×256 random patches, two augmented samples per minibatch, and Adam with a scheduled learning-rate decay. PUCA uses a 3-level U-Net with DAB blocks, selects stride factors to disrupt noise correlations while preserving global semantic structure, and applies random-replacing refinement post-processing. The receptive-field visualization computes gradient-based influence scores from the output central pixel, scaled and clipped for comparison across levels.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/puca-patch-unshuffle-and-channel-attention-for-enhanced-self-supervised-image-denoising-implementation-details/513532/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/puca-patch-unshuffle-and-channel-attention-for-enhanced-self-supervised-image-denoising-implementation-details/513532.png","ImageObject",300,407,{"name":92,"@type":93},"dawugda","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-16","2026-10-06",true,{"@type":102,"interactionType":103,"userInteractionCount":39},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What datasets and training patch settings are used for PUCA?","Question",{"text":112,"@type":113},"PUCA trains on SIDD-Medium and DND using randomly cropped 256×256 patches. Each epoch uses 24,542 patches from SIDD-Medium and 24,784 patches from DND, with random 90° rotation and horizontal/vertical flips for augmentation.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"What network architecture does PUCA use and how are stride factors selected?",{"text":117,"@type":113},"PUCA uses a 3-level U-Net with L1=3, L2=4, and L3=4 DAB blocks per level. It selects a stride factor of 4 during training and uses stride factor 2 at test time to break noise correlations while limiting global semantic degradation.",{"name":119,"@type":110,"acceptedAnswer":120},"How are receptive fields visualized in the supplemental material?",{"text":121,"@type":113},"Receptive fields are visualized by computing the influence of an input pixel on the output central pixel’s summed channel score. The method takes the absolute gradient with respect to the input pixel, then uses the maximum over channels; values are scaled by 50 and clipped to [0,3] for level-wise comparison.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},513532,1791368319,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":39,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":19,"language":139,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":140,"faqs":141,"seo_title":142,"seo_description":67,"update_tm":143,"read_time":52,"preview_page_count":19,"preview_urls":144},3985747870947,"https://ap-avatar.wpscdn.com/davatar_3d24733baf745e90a7e4bdd5f77d97b2","Supplementary Material  \nPUCA: Patch-Unshuffle and Channel Attention for Enhanced Self-Supervised Image Denoising  \nHyemi Jang 1 , Junsung Park 1 , Dahuin Jung 1 , Jaihyun Lew2 , Ho Bae3 ,  Sungroh Yoon 1 ,2 , ∗  \n1Department of Electrical and Computer Engineering, Seoul National University  \n2Interdisciplinary Program in Artificial Intelligence, Seoul National University  \n3Department of Cyber Security, Ewha Womans University  \n[wkdal9512@snu.ac.kr](wkdal9512@snu.ac.kr) [jerryray@snu.ac.kr](jerryray@snu.ac.kr) [annajung0625@snu.ac.kr](annajung0625@snu.ac.kr)[ ](annajung0625@snu.ac.kr)[fudojhl@snu.ac.kr](fudojhl@snu.ac.kr) [hobae@ewha.ac.kr](hobae@ewha.ac.kr) [sryoon@snu.ac.kr](sryoon@snu.ac.kr)  \n1 Implementation details  \n1.1 Training  \nWe use the SIDD-Medium and DND datasets for training. For SIDD validation and SIDD benchmark evaluation, we utilize a model trained on the SIDD-Medium dataset, while for DND benchmark evaluation, we use a model trained on the DND benchmark dataset. To train our PUCA, we employ randomly cropped patches of size 256 × 256 from the SIDD-Medium and DND datasets. Each epoch involves 24,542 patches from the SIDD-Medium dataset and 24,784 patches from the DND dataset. To enhance the diversity of our training samples, each sample is augmented with a random 90◦ rotation and horizontal/vertical flips. Our minibatch consists of two augmented samples. The learning rate of the Adam optimizer is reduced by a factor of 10 every 8 epochs. The SIDD training took approximately 44.38 hours, while the DND training took approximately 44.78 hours.  \n1.2 Network  \nPUCA has a 3-level U-Net structure, with L 1 =3, L2 =4, and L3 =4 DAB blocks at each level. From the observations of AP-BSN [3], it was confirmed that comparable denoising results were obtained when the stride factor of pixel-shuffle downsampling had a value between 4 and 6 during training. To destroy noise correlations while minimizing degradation of the global semantic structure due to the combined effects of pixel-shuffle downsampling and patch-unshuffle, we select a stride factor of 4 . The stride factor of test time is 2 . For the post-processing strategy, we adopt the random-replacing refinement (R3 ) method introduced in Lee et al. [3] and set T = 8 and p = 0 .16. We implement PUCA based on the AP-BSN code ([https://github.com/wooseoklee4/AP-BSN.git](https://github.com/wooseoklee4/AP-BSN.git)).  \n1.3 Receptive Field Visualization  \nIn this section, we explain the implementation of the receptive fields visualized in Figure 5 of the main text. Similar to a saliency map, we calculate the influence that the input pixel has on the central pixel of the output. We use the sum of the channels of the output’s central pixel as the score and compute the gradient with respect to the input pixel of the score. The gradient indicates how much the score changes when the pixel undergoes a small modification. We take the absolute value of the gradient and then calculate the influence by finding the maximum value along the channel axis. To accentuate the differences between the receptive fields of AP-BSN and each level, we scale the calculated values by a factor of 50 and clip them to a range between 0 and 3 .  \n∗ Corresponding Authors  \n37th Conference on Neural Information Processing Systems (NeurIPS 2023) .  \n2 Additional Qualitative Results  \nWe visualize the denoising results for additional samples. The baseline models used for comparison are as follows, and the samples from these baseline models are generated from the checkpoints they provide.  \nZhou et al. [6] trains a denoising model from additive white Gaussian noise and applies the model to real noisy images where the noise correlation is broken by pixel-shuffle downsampling with astride of 2 . C2N trains a generator to create pseudo-noisy images and then trains a denoising model like DIDN using noisy-clean pairs (C2N+DIDN) [2] . R2R [4] trains the denoising models using noisy pairs created from unorganized ","cbCaippKFpw4weSJ","https://ap.wps.com/l/cbCaippKFpw4weSJ","pdf",13124428,"English","# 1 Implementation details\n## 1.1 Training\n## 1.2 Network\n## 1.3 Receptive Field Visualization\n# 2 Additional Qualitative Results\n# References","[{\"question\":\"What datasets and training patch settings are used for PUCA?\",\"answer\":\"PUCA trains on SIDD-Medium and DND using randomly cropped 256×256 patches. Each epoch uses 24,542 patches from SIDD-Medium and 24,784 patches from DND, with random 90° rotation and horizontal/vertical flips for augmentation.\"},{\"question\":\"What network architecture does PUCA use and how are stride factors selected?\",\"answer\":\"PUCA uses a 3-level U-Net with L1=3, L2=4, and L3=4 DAB blocks per level. It selects a stride factor of 4 during training and uses stride factor 2 at test time to break noise correlations while limiting global semantic degradation.\"},{\"question\":\"How are receptive fields visualized in the supplemental material?\",\"answer\":\"Receptive fields are visualized by computing the influence of an input pixel on the output central pixel’s summed channel score. The method takes the absolute gradient with respect to the input pixel, then uses the maximum over channels; values are scaled by 50 and clipped to [0,3] for level-wise comparison.\"}]","PUCA: Patch-Unshuffle and Channel Attention for Enhanced Self-Supervised Image Denoising - Implementation Details | PDF",1791286745,[145,149,153,157],{"page_no":8,"url":146,"width":147,"height":148},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/93c3e1959e967216cab03bbaa48adfb6.webp",983,1272,{"page_no":14,"url":150,"width":151,"height":152},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/ca94021ad5351ffe3f85be6ca329029c.webp",982,1270,{"page_no":81,"url":154,"width":155,"height":156},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/a5f2d48b2904be5c68c692f45b91676b.webp",981,1269,{"page_no":19,"url":158,"width":159,"height":160},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/9b5b732444535d6ac0c1944b95574513.webp",980,1268]