<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[fantasy_ARM9的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/fantasy_ARM9</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; fantasy_ARM9]]></copyright><item><title><![CDATA[BasicVSR-lite图像画质增强]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161455202</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161455202</guid><author>fantasy_ARM9</author><pubDate>Wed, 27 May 2026 14:51:35 +0800</pubDate><description><![CDATA[BasicVSR++论文也明确说，基于recurrent structure 通过双向传播和特征对其来利用整个视频序列的信息，BasicVSR++进一步加入二阶传播和flow-guided deformable alignment，增强了对错位视频帧的时空信息利用。BasicVSR类模型通常使用光流网络，比如SPyNet， 来估计相邻帧之间的运动，BasicVSR++补充材料里也提到使用pretrained SPyNet作为flow network。低清视频帧，[B, 7, 3, 64, 64]]]></description><category></category></item><item><title><![CDATA[pytorch人脸匹配模型]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161291009</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161291009</guid><author>fantasy_ARM9</author><pubDate>Thu, 21 May 2026 18:20:54 +0800</pubDate><description><![CDATA[conv1   3x3卷积，64通道,stride = 1， 输出 [B, 64, 112x112]layer3 ResNet BasicBlock x2 下采样， [B, 256, 28, 28]1 人脸检查，人脸裁剪，人脸关键点检测，根据双眼位置做人脸对齐，保存成112x112的人脸图。输入 RGB人脸图像， 输出尺寸 [B,3,112,112]avgpool 全局平均池化， [B, 512, 1, 1]组成batch [B,3,112,112] 几张图一组。]]></description><category></category></item><item><title><![CDATA[fpga demo测试hello world]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161116738</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/161116738</guid><author>fantasy_ARM9</author><pubDate>Fri, 15 May 2026 15:45:14 +0800</pubDate><description><![CDATA[二 Hello world demo代码。一 看下机器上是什么fpga设备。]]></description><category></category></item><item><title><![CDATA[SVT-AV1 整体架构]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159964482</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159964482</guid><author>fantasy_ARM9</author><pubDate>Wed, 08 Apr 2026 19:22:22 +0800</pubDate><description><![CDATA[>Picture Decision svt_aom_picture_decision_kernel 显示序重排， + GOP类型。->EncDec / Mode Decision svt_aom_mode_decision_kernel, 按SB波前多线程。->Picture Analysis (svt_aom_picture_analysis_kernel, 可以多线程)->Picture Manager svt_aom_picture_manager_kernel, 子图调度，参考关系。]]></description><category></category></item><item><title><![CDATA[SVT-AV1 维纳滤波的使用]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159950265</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159950265</guid><author>fantasy_ARM9</author><pubDate>Wed, 08 Apr 2026 10:28:42 +0800</pubDate><description><![CDATA[apply_denoise_2d 函数分析。]]></description><category></category></item><item><title><![CDATA[X265墒编码--代码分析]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159432929</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/159432929</guid><author>fantasy_ARM9</author><pubDate>Tue, 24 Mar 2026 18:13:43 +0800</pubDate><description><![CDATA[x265 墒编码X265 HEVC编码器架构分析一 整体代码架构1.1 目录与模块划分source/├── x265cli.cpp / x265cli.h    # 命令行入口、参数解析、help├── x265.h                      # 对外 API、参数结构、版本├── encoder/                    # 编码核心│   ├── encoder.cpp/h          # 编码器顶层：创建/销毁、encode() 主流程、VPS/SPS/PPS│  ]]></description><category></category></item><item><title><![CDATA[SVT-AV1四叉树划分-模式决策]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157549227</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157549227</guid><author>fantasy_ARM9</author><pubDate>Fri, 30 Jan 2026 15:26:46 +0800</pubDate><description><![CDATA[AV1标准包含矩形划分，PARTITION_HORZ, PARTITION_VERT, PARTITION_HORZ_4等，而set_child_to_be_considered 仅仅处理四叉树，PARTITION_SPLIT 的递归。由块几何表，svt_aom_blk_geom_mds 和d1 阶段决策如 svt_aom_d1_non_square_block_decision。矩形划分的cost 比较在d1阶段，例如svt_aom_d1_non_square_block_decision。]]></description><category></category></item><item><title><![CDATA[SVT-AV1编码 递归子块划分]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157515534</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157515534</guid><author>fantasy_ARM9</author><pubDate>Thu, 29 Jan 2026 19:17:40 +0800</pubDate><description><![CDATA[当一个块需要继续划分时（split_flag=TRUE），本函数会递归地将其4个子块。set_child_to_be_considered() 递归标记哪些块需要处理。* 标记为需要处理，如果子块也需要继续划分，则继续递归标记子块的子块。这是AV1编码中实现多叉树（四叉树）递归划分的第一阶段：递归标记。本函数只负责递归标记哪些块需要处理，并不计算COST。* - 64x64 -> 4个 32x32。* - 32x32 -> 4个 16x16。* - 8x8 -> 4个 4x4。]]></description><category></category></item><item><title><![CDATA[SVT-AV1 md_stage_1 函数分析]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157256896</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157256896</guid><author>fantasy_ARM9</author><pubDate>Thu, 22 Jan 2026 16:36:36 +0800</pubDate><description><![CDATA[【代码】SVT-AV1 md_stage_1 函数分析。]]></description><category></category></item><item><title><![CDATA[SVT-AV1中的global-MV/仿射运动实现分析]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157225064</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/157225064</guid><author>fantasy_ARM9</author><pubDate>Thu, 22 Jan 2026 00:00:22 +0800</pubDate><description><![CDATA[这是一个路由函数，router function，根据配置的 correspondence_method, 选择使用哪种方法来收集对应点，correspondences, 对应点是全局运动估计的基础数据。MV方法，(MV_64x64, MV_32x32, MV_16x16/MV_8x8)2 合并阶段，RANSAC算法，将收集到的对应点聚合成全局运动模型。原理：基于图像特征，角点进行匹配，不依赖运动估计结果。优点：更鲁棒，精度高，不依赖ME质量，适合复杂运动。适用场景：复杂运动，ME质量差时，需要高精度时。]]></description><category></category></item><item><title><![CDATA[AV1全局运动补偿]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156956723</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156956723</guid><author>fantasy_ARM9</author><pubDate>Thu, 15 Jan 2026 00:18:05 +0800</pubDate><description><![CDATA[处理单参考帧GLOBALMV和双参考帧 GLOBAL_GLOBALMV 模式。determine_gm_params() 确定全局运动参数。inject_global_candidates()函数。扭曲运动（Warped Motion）相关实现。RANSAC 算法实现（用于全局运动参数估计）将全局运动候选注入到候选数组。角点检测（用于全局运动估计）角点匹配（用于全局运动估计）全局运动估计的成本计算相关。全局运动使用扭曲运动模型。]]></description><category></category></item><item><title><![CDATA[SVT-AV1 模式决策函数md_encode_block]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156952114</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156952114</guid><author>fantasy_ARM9</author><pubDate>Wed, 14 Jan 2026 19:43:00 +0800</pubDate><description><![CDATA[inject_intra_candidates() 注入帧内预测候选，DC，方向模式，角度delta。inject_filter_intra_candidates() 注入滤波帧内候选。inject_intra_bc_candidates() 注入帧内复制IBC 候选。svt_aom_inject_inter_candidates() 注入帧间候选。阶段2 候选生成 gernerate_md_stage_0_cand。4 帧间相关，运动估计，ME和预测运动估计PME。双参考帧模式，双向预测。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion -- DDPM]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156774739</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156774739</guid><author>fantasy_ARM9</author><pubDate>Fri, 09 Jan 2026 23:58:16 +0800</pubDate><description><![CDATA[【代码】Stable Diffusion -- DDPM。]]></description><category></category></item><item><title><![CDATA[Transformer vs Stable Diffusion vs LLM模型对比]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156737065</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156737065</guid><author>fantasy_ARM9</author><pubDate>Fri, 09 Jan 2026 00:02:56 +0800</pubDate><description><![CDATA[三 Stable Diffusion模型实现示例。二 下面我们详细拆解。]]></description><category></category></item><item><title><![CDATA[AlexNet分析和代码实现]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156714206</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156714206</guid><author>fantasy_ARM9</author><pubDate>Thu, 08 Jan 2026 10:13:51 +0800</pubDate><description><![CDATA[二 AlexNet的网络架构。三 AlexNet创新点。一 AlexNet特点。]]></description><category></category></item><item><title><![CDATA[U-Net图像分割模型]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156659565</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156659565</guid><author>fantasy_ARM9</author><pubDate>Wed, 07 Jan 2026 00:09:28 +0800</pubDate><description><![CDATA[试图找到潜在的癌性病变，图像分割发挥作用的地方，分割过程用于识别图像中的不同组织和结构，在区分癌细胞和其他正常组织方面发挥着重要作用。U-net 的命名是因为结构类似于字母U，我们在输出端得到分割后的输入图像，U-net的架构师独特的，因为它由收缩路径和扩展路径组成。通常在像素级别进行，通过将图像中具有相似特征的区域分组或定义对象的边界来完成，这是一种识别和解析图像中不同对象或特征的方法。尽管U-Net专注于生物医学图像，其灵活的结构允许它有效的用于其他类型的图像数据。]]></description><category></category></item><item><title><![CDATA[SVT-AV1帧内编码代码分析]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156507288</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156507288</guid><author>fantasy_ARM9</author><pubDate>Fri, 02 Jan 2026 10:03:41 +0800</pubDate><description><![CDATA[svt_aom_mode_decision_sb 解析。]]></description><category></category></item><item><title><![CDATA[AV1视频编码位于图像边界的超级块划分]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156434005</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156434005</guid><author>fantasy_ARM9</author><pubDate>Wed, 31 Dec 2025 00:05:07 +0800</pubDate><description><![CDATA[一 位于图像边界的超级块划分。]]></description><category></category></item><item><title><![CDATA[ResNet网络模型]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156401138</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156401138</guid><author>fantasy_ARM9</author><pubDate>Tue, 30 Dec 2025 00:25:27 +0800</pubDate><description><![CDATA[【代码】ResNet网络模型。]]></description><category></category></item><item><title><![CDATA[SVT-AV1 B帧决策和mini-GOP决策分析]]></title><link>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156395694</link><guid>https://blog.youkuaiyun.com/fantasy_ARM9/article/details/156395694</guid><author>fantasy_ARM9</author><pubDate>Mon, 29 Dec 2025 19:16:19 +0800</pubDate><description><![CDATA[三 核心决策函数 calc_mini_gop_activity。二 mini-GOP的决策。]]></description><category></category></item></channel></rss>