You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

README.md 21 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500
  1. ## MindSpore Lite 端侧目标检测demo(Android)
  2. 本示例程序演示了如何在端侧利用MindSpore Lite C++ API(Android JNI)以及MindSpore Lite 目标检测模型完成端侧推理,实现对图库或者设备摄像头捕获的内容进行检测,并在App图像预览界面中显示连续目标检测结果。
  3. ### 运行依赖
  4. - Android Studio >= 3.2 (推荐4.0以上版本)
  5. - NDK 21.3
  6. - CMake 3.10
  7. - Android SDK >= 26
  8. ### 构建与运行
  9. 1. 在Android Studio中加载本示例源码,并安装相应的SDK(指定SDK版本后,由Android Studio自动安装)。
  10. ![start_home](images/home.png)
  11. 启动Android Studio后,点击`File->Settings->System Settings->Android SDK`,勾选相应的SDK。如下图所示,勾选后,点击`OK`,Android Studio即可自动安装SDK。
  12. ![start_sdk](images/sdk_management.png)
  13. (可选)若安装时出现NDK版本问题,可手动下载相应的[NDK版本](https://developer.android.com/ndk/downloads?hl=zh-cn)(本示例代码使用的NDK版本为21.3),并在`Project Structure`的`Android NDK location`设置中指定SDK的位置。
  14. ![project_structure](images/project_structure.png)
  15. 2. 连接Android设备,运行目标检测示例应用程序。
  16. 通过USB连接Android设备调试,点击`Run 'app'`即可在你的设备上运行本示例项目。
  17. * 注:编译过程中Android Studio会自动下载MindSpore Lite、模型文件等相关依赖项,编译过程需做耐心等待。
  18. ![run_app](images/run_app.PNG)
  19. Android Studio连接设备调试操作,可参考<https://developer.android.com/studio/run/device?hl=zh-cn>。
  20. 3. 在Android设备上,点击“继续安装”,安装完即可查看到设备摄像头捕获的内容和推理结果。
  21. ![install](images/install.jpg)
  22. 如下图所示,检测出图中内容是鼠标。
  23. ![result](images/object_detection.png)
  24. ## 示例程序详细说明
  25. 本端侧目标检测Android示例程序分为JAVA层和JNI层,其中,JAVA层主要通过Android Camera 2 API实现摄像头获取图像帧,以及相应的图像处理(针对推理结果画框)等功能;JNI层在[Runtime](https://www.mindspore.cn/tutorial/lite/zh-CN/master/use/runtime.html)中完成模型推理的过程。
  26. > 此处详细说明示例程序的JNI层实现,JAVA层运用Android Camera 2 API实现开启设备摄像头以及图像帧处理等功能,需读者具备一定的Android开发基础知识。
  27. ### 示例程序结构
  28. ```
  29. app
  30. |
  31. ├── libs # 存放demo jni层编译出的库文件
  32. │ └── arm64-v8a
  33. │ │── libmlkit-label-MS.so #
  34. |
  35. ├── src/main
  36. │ ├── assets # 资源文件
  37. | | └── ssd.ms # 存放模型文件
  38. │ |
  39. │ ├── cpp # 模型加载和预测主要逻辑封装类
  40. | | ├── mindspore-lite-x.x.x-mindata-arm64-cpu # minspore源码编译出的调用包,包含demo jni层依赖的库文件及相关的头文件
  41. | | | └── ...
  42. │ | |
  43. | | ├── MindSporeNetnative.cpp # MindSpore调用相关的JNI方法
  44. │ ├── java # java层应用代码
  45. │ │ └── com.huawei.himindsporedemo
  46. │ │ ├── help # 图像处理及MindSpore JNI调用相关实现
  47. │ │ │ └── ...
  48. │ │ └── obejctdetect # 开启摄像头及绘制相关实现
  49. │ │ └── ...
  50. │ │
  51. │ ├── res # 存放Android相关的资源文件
  52. │ └── AndroidManifest.xml # Android配置文件
  53. ├── CMakeList.txt # cmake编译入口文件
  54. ├── build.gradle # 其他Android配置文件
  55. ├── download.gradle # APP构建时由gradle自动从HuaWei Server下载依赖的库文件及模型文件
  56. └── ...
  57. ```
  58. ### 配置MindSpore Lite依赖项
  59. Android JNI层调用MindSpore C++ API时,需要相关库文件支持。可通过MindSpore Lite[源码编译](https://www.mindspore.cn/tutorial/lite/zh-CN/master/use/build.html)生成`mindspore-lite-{version}-minddata-{os}-{device}.tar.gz`库文件包并解压缩(包含`libmindspore-lite.so`库文件和相关头文件),在本例中需使用生成带图像预处理模块的编译命令。
  60. > version:输出件版本号,与所编译的分支代码对应的版本一致。
  61. >
  62. > device:当前分为cpu(内置CPU算子)和gpu(内置CPU和GPU算子)。
  63. >
  64. > os:输出件应部署的操作系统。
  65. 本示例中,build过程由download.gradle文件自动下载MindSpore Lite 版本文件,并放置在`app/src/main/cpp/`目录下。
  66. * 注:若自动下载失败,请手动下载相关库文件,解压并放在对应位置:
  67. mindspore-lite-1.0.1-runtime-arm64-cpu.tar.gz [下载链接](https://ms-release.obs.cn-north-4.myhuaweicloud.com/1.0.1/lite/android_aarch64/mindspore-lite-1.0.1-runtime-arm64-cpu.tar.gz)
  68. 在app的`build.gradle`文件中配置CMake编译支持,以及`arm64-v8a`的编译支持,如下所示:
  69. ```
  70. android{
  71. defaultConfig{
  72. externalNativeBuild{
  73. cmake{
  74. arguments "-DANDROID_STL=c++_shared"
  75. }
  76. }
  77. ndk{
  78. abiFilters 'arm64-v8a'
  79. }
  80. }
  81. }
  82. ```
  83. 在`app/CMakeLists.txt`文件中建立`.so`库文件链接,如下所示。
  84. ```
  85. # Set MindSpore Lite Dependencies.
  86. set(MINDSPORELITE_VERSION mindspore-lite-1.0.1-runtime-arm64-cpu)
  87. include_directories(${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION})
  88. add_library(mindspore-lite SHARED IMPORTED )
  89. add_library(minddata-lite SHARED IMPORTED )
  90. set_target_properties(mindspore-lite PROPERTIES IMPORTED_LOCATION
  91. ${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION}/lib/libmindspore-lite.so)
  92. set_target_properties(minddata-lite PROPERTIES IMPORTED_LOCATION
  93. ${CMAKE_SOURCE_DIR}/src/main/cpp/${MINDSPORELITE_VERSION}/lib/libminddata-lite.so)
  94. # Link target library.
  95. target_link_libraries(
  96. ...
  97. mindspore-lite
  98. minddata-lite
  99. ...
  100. )
  101. ```
  102. ### 下载及部署模型文件
  103. 从MindSpore Model Hub中下载模型文件,本示例程序中使用的目标检测模型文件为`ssd.ms`,同样通过`download.gradle`脚本在APP构建时自动下载,并放置在`app/src/main/assets`工程目录下。
  104. * 注:若下载失败请手动下载模型文件,ssd.ms [下载链接](https://download.mindspore.cn/model_zoo/official/lite/ssd_mobilenetv2_lite/ssd.ms)。
  105. ### 编写端侧推理代码
  106. 在JNI层调用MindSpore Lite C++ API实现端测推理。
  107. 推理代码流程如下,完整代码请参见`src/cpp/MindSporeNetnative.cpp`。
  108. 1. 加载MindSpore Lite模型文件,构建上下文、会话以及用于推理的计算图。
  109. - 加载模型文件:创建并配置用于模型推理的上下文
  110. ```cpp
  111. // Buffer is the model data passed in by the Java layer
  112. jlong bufferLen = env->GetDirectBufferCapacity(buffer);
  113. char *modelBuffer = CreateLocalModelBuffer(env, buffer);
  114. ```
  115. - 创建会话
  116. ```cpp
  117. void **labelEnv = new void *;
  118. MSNetWork *labelNet = new MSNetWork;
  119. *labelEnv = labelNet;
  120. // Create context.
  121. lite::Context *context = new lite::Context;
  122. context->cpu_bind_mode_ = lite::NO_BIND;
  123. context->device_ctx_.type = lite::DT_CPU;
  124. context->thread_num_ = numThread; //Specify the number of threads to run inference
  125. // Create the mindspore session.
  126. labelNet->CreateSessionMS(modelBuffer, bufferLen, "device label", context);
  127. delete context;
  128. ```
  129. - 加载模型文件并构建用于推理的计算图
  130. ```cpp
  131. void MSNetWork::CreateSessionMS(char* modelBuffer, size_t bufferLen, std::string name, mindspore::lite::Context* ctx)
  132. {
  133. CreateSession(modelBuffer, bufferLen, ctx);
  134. session = mindspore::session::LiteSession::CreateSession(ctx);
  135. auto model = mindspore::lite::Model::Import(modelBuffer, bufferLen);
  136. int ret = session->CompileGraph(model); // Compile Graph
  137. }
  138. ```
  139. 2. 将输入图片转换为传入MindSpore模型的Tensor格式。
  140. 将待检测图片数据转换为输入MindSpore模型的Tensor。
  141. ```cpp
  142. // Convert the Bitmap image passed in from the JAVA layer to Mat for OpenCV processing
  143. LiteMat lite_mat_bgr,lite_norm_mat_cut;
  144. if (!BitmapToLiteMat(env, srcBitmap, lite_mat_bgr)){
  145. MS_PRINT("BitmapToLiteMat error");
  146. return NULL;
  147. }
  148. int srcImageWidth = lite_mat_bgr.width_;
  149. int srcImageHeight = lite_mat_bgr.height_;
  150. if(!PreProcessImageData(lite_mat_bgr, lite_norm_mat_cut)){
  151. MS_PRINT("PreProcessImageData error");
  152. return NULL;
  153. }
  154. ImgDims inputDims;
  155. inputDims.channel =lite_norm_mat_cut.channel_;
  156. inputDims.width = lite_norm_mat_cut.width_;
  157. inputDims.height = lite_norm_mat_cut.height_;
  158. // Get the mindsore inference environment which created in loadModel().
  159. void **labelEnv = reinterpret_cast<void **>(netEnv);
  160. if (labelEnv == nullptr) {
  161. MS_PRINT("MindSpore error, labelEnv is a nullptr.");
  162. return NULL;
  163. }
  164. MSNetWork *labelNet = static_cast<MSNetWork *>(*labelEnv);
  165. auto mSession = labelNet->session;
  166. if (mSession == nullptr) {
  167. MS_PRINT("MindSpore error, Session is a nullptr.");
  168. return NULL;
  169. }
  170. MS_PRINT("MindSpore get session.");
  171. auto msInputs = mSession->GetInputs();
  172. auto inTensor = msInputs.front();
  173. float *dataHWC = reinterpret_cast<float *>(lite_norm_mat_cut.data_ptr_);
  174. // copy input Tensor
  175. memcpy(inTensor->MutableData(), dataHWC,
  176. inputDims.channel * inputDims.width * inputDims.height * sizeof(float));
  177. delete[] (dataHWC);
  178. ```
  179. 3. 进行模型推理前,输入tensor格式为 NHWC,shape为1:300:300:3,格式为RGB, 并对输入tensor做标准化处理.
  180. ```cpp
  181. bool PreProcessImageData(LiteMat &lite_mat_bgr,LiteMat &lite_norm_mat_cut) {
  182. bool ret=false;
  183. LiteMat lite_mat_resize;
  184. ret = ResizeBilinear(lite_mat_bgr, lite_mat_resize, 300, 300);
  185. if (!ret) {
  186. MS_PRINT("ResizeBilinear error");
  187. return false;
  188. }
  189. LiteMat lite_mat_convert_float;
  190. ret = ConvertTo(lite_mat_resize, lite_mat_convert_float, 1.0 / 255.0);
  191. if (!ret) {
  192. MS_PRINT("ConvertTo error");
  193. return false;
  194. }
  195. float means[3] = {0.485, 0.456, 0.406};
  196. float vars[3] = {1.0 / 0.229, 1.0 / 0.224, 1.0 / 0.225};
  197. SubStractMeanNormalize(lite_mat_convert_float, lite_norm_mat_cut, means, vars);
  198. return true;
  199. }
  200. ```
  201. 4. 对输入Tensor按照模型进行推理,获取输出Tensor,并进行后处理。
  202. - 图执行,端测推理。
  203. ```cpp
  204. // After the model and image tensor data is loaded, run inference.
  205. auto status = mSession->RunGraph();
  206. ```
  207. - 获取输出数据。
  208. ```cpp
  209. auto names = mSession->GetOutputTensorNames();
  210. typedef std::unordered_map<std::string,
  211. std::vector<mindspore::tensor::MSTensor *>> Msout;
  212. std::unordered_map<std::string,
  213. mindspore::tensor::MSTensor *> msOutputs;
  214. for (const auto &name : names) {
  215. auto temp_dat =mSession->GetOutputByTensorName(name);
  216. msOutputs.insert(std::pair<std::string, mindspore::tensor::MSTensor *> {name, temp_dat});
  217. }
  218. std::string retStr = ProcessRunnetResult(msOutputs, ret);
  219. ```
  220. - 模型有2个输出,输出1是目标的类别置信度,维度为1:1917: 81; 输出2是目标的矩形框坐标偏移量,维度为1:1917:4。 为了得出目标的实际矩形框,需要根据偏移量计算出矩形框的位置。这部分在 getDefaultBoxes中实现。
  221. ```cpp
  222. void SSDModelUtil::getDefaultBoxes() {
  223. float fk[6] = {0.0, 0.0, 0.0, 0.0, 0.0, 0.0};
  224. std::vector<struct WHBox> all_sizes;
  225. struct Product mProductData[19 * 19] = {0};
  226. for (int i = 0; i < 6; i++) {
  227. fk[i] = config.model_input_height / config.steps[i];
  228. }
  229. float scale_rate =
  230. (config.max_scale - config.min_scale) / (sizeof(config.num_default) / sizeof(int) - 1);
  231. float scales[7] = {0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0};
  232. for (int i = 0; i < sizeof(config.num_default) / sizeof(int); i++) {
  233. scales[i] = config.min_scale + scale_rate * i;
  234. }
  235. for (int idex = 0; idex < sizeof(config.feature_size) / sizeof(int); idex++) {
  236. float sk1 = scales[idex];
  237. float sk2 = scales[idex + 1];
  238. float sk3 = sqrt(sk1 * sk2);
  239. struct WHBox tempWHBox;
  240. all_sizes.clear();
  241. if (idex == 0) {
  242. float w = sk1 * sqrt(2);
  243. float h = sk1 / sqrt(2);
  244. tempWHBox.boxw = 0.1;
  245. tempWHBox.boxh = 0.1;
  246. all_sizes.push_back(tempWHBox);
  247. tempWHBox.boxw = w;
  248. tempWHBox.boxh = h;
  249. all_sizes.push_back(tempWHBox);
  250. tempWHBox.boxw = h;
  251. tempWHBox.boxh = w;
  252. all_sizes.push_back(tempWHBox);
  253. } else {
  254. tempWHBox.boxw = sk1;
  255. tempWHBox.boxh = sk1;
  256. all_sizes.push_back(tempWHBox);
  257. for (int j = 0; j < sizeof(config.aspect_ratios[idex]) / sizeof(int); j++) {
  258. float w = sk1 * sqrt(config.aspect_ratios[idex][j]);
  259. float h = sk1 / sqrt(config.aspect_ratios[idex][j]);
  260. tempWHBox.boxw = w;
  261. tempWHBox.boxh = h;
  262. all_sizes.push_back(tempWHBox);
  263. tempWHBox.boxw = h;
  264. tempWHBox.boxh = w;
  265. all_sizes.push_back(tempWHBox);
  266. }
  267. tempWHBox.boxw = sk3;
  268. tempWHBox.boxh = sk3;
  269. all_sizes.push_back(tempWHBox);
  270. }
  271. for (int i = 0; i < config.feature_size[idex]; i++) {
  272. for (int j = 0; j < config.feature_size[idex]; j++) {
  273. mProductData[i * config.feature_size[idex] + j].x = i;
  274. mProductData[i * config.feature_size[idex] + j].y = j;
  275. }
  276. }
  277. int productLen = config.feature_size[idex] * config.feature_size[idex];
  278. for (int i = 0; i < productLen; i++) {
  279. for (int j = 0; j < all_sizes.size(); j++) {
  280. struct NormalBox tempBox;
  281. float cx = (mProductData[i].y + 0.5) / fk[idex];
  282. float cy = (mProductData[i].x + 0.5) / fk[idex];
  283. tempBox.y = cy;
  284. tempBox.x = cx;
  285. tempBox.h = all_sizes[j].boxh;
  286. tempBox.w = all_sizes[j].boxw;
  287. mDefaultBoxes.push_back(tempBox);
  288. }
  289. }
  290. }
  291. }
  292. ```
  293. - 通过最大值抑制将目标类型置信度较高的输出筛选出来。
  294. ```cpp
  295. void SSDModelUtil::nonMaximumSuppression(const YXBoxes *const decoded_boxes,
  296. const float *const scores,
  297. const std::vector<int> &in_indexes,
  298. std::vector<int> &out_indexes, const float nmsThreshold,
  299. const int count, const int max_results) {
  300. int nR = 0; //number of results
  301. std::vector<bool> del(count, false);
  302. for (size_t i = 0; i < in_indexes.size(); i++) {
  303. if (!del[in_indexes[i]]) {
  304. out_indexes.push_back(in_indexes[i]);
  305. if (++nR == max_results) {
  306. break;
  307. }
  308. for (size_t j = i + 1; j < in_indexes.size(); j++) {
  309. const auto boxi = decoded_boxes[in_indexes[i]], boxj = decoded_boxes[in_indexes[j]];
  310. float a[4] = {boxi.xmin, boxi.ymin, boxi.xmax, boxi.ymax};
  311. float b[4] = {boxj.xmin, boxj.ymin, boxj.xmax, boxj.ymax};
  312. if (IOU(a, b) > nmsThreshold) {
  313. del[in_indexes[j]] = true;
  314. }
  315. }
  316. }
  317. }
  318. }
  319. ```
  320. - 对每类的概率大于阈值,通过NMS算法筛选出矩形框后, 还需要将输出的矩形框恢复到原图尺寸。
  321. ```cpp
  322. std::string SSDModelUtil::getDecodeResult(float *branchScores, float *branchBoxData) {
  323. std::string result = "";
  324. NormalBox tmpBox[1917] = {0};
  325. float mScores[1917][81] = {0};
  326. float outBuff[1917][7] = {0};
  327. float scoreWithOneClass[1917] = {0};
  328. int outBoxNum = 0;
  329. YXBoxes decodedBoxes[1917] = {0};
  330. // Copy branch outputs box data to tmpBox.
  331. for (int i = 0; i < 1917; ++i) {
  332. tmpBox[i].y = branchBoxData[i * 4 + 0];
  333. tmpBox[i].x = branchBoxData[i * 4 + 1];
  334. tmpBox[i].h = branchBoxData[i * 4 + 2];
  335. tmpBox[i].w = branchBoxData[i * 4 + 3];
  336. }
  337. // Copy branch outputs score to mScores.
  338. for (int i = 0; i < 1917; ++i) {
  339. for (int j = 0; j < 81; ++j) {
  340. mScores[i][j] = branchScores[i * 81 + j];
  341. }
  342. }
  343. ssd_boxes_decode(tmpBox, decodedBoxes);
  344. const float nms_threshold = 0.3;
  345. for (int i = 1; i < 81; i++) {
  346. std::vector<int> in_indexes;
  347. for (int j = 0; j < 1917; j++) {
  348. scoreWithOneClass[j] = mScores[j][i];
  349. // if (mScores[j][i] > 0.1) {
  350. if (mScores[j][i] > g_thres_map[i]) {
  351. in_indexes.push_back(j);
  352. }
  353. }
  354. if (in_indexes.size() == 0) {
  355. continue;
  356. }
  357. sort(in_indexes.begin(), in_indexes.end(),
  358. [&](int a, int b) { return scoreWithOneClass[a] > scoreWithOneClass[b]; });
  359. std::vector<int> out_indexes;
  360. nonMaximumSuppression(decodedBoxes, scoreWithOneClass, in_indexes, out_indexes,
  361. nms_threshold);
  362. for (int k = 0; k < out_indexes.size(); k++) {
  363. outBuff[outBoxNum][0] = out_indexes[k]; //image id
  364. outBuff[outBoxNum][1] = i; //labelid
  365. outBuff[outBoxNum][2] = scoreWithOneClass[out_indexes[k]]; //scores
  366. outBuff[outBoxNum][3] =
  367. decodedBoxes[out_indexes[k]].xmin * inputImageWidth / 300;
  368. outBuff[outBoxNum][4] =
  369. decodedBoxes[out_indexes[k]].ymin * inputImageHeight / 300;
  370. outBuff[outBoxNum][5] =
  371. decodedBoxes[out_indexes[k]].xmax * inputImageWidth / 300;
  372. outBuff[outBoxNum][6] =
  373. decodedBoxes[out_indexes[k]].ymax * inputImageHeight / 300;
  374. outBoxNum++;
  375. }
  376. }
  377. MS_PRINT("outBoxNum %d", outBoxNum);
  378. for (int i = 0; i < outBoxNum; ++i) {
  379. std::string tmpid_str = std::to_string(outBuff[i][0]);
  380. result += tmpid_str; // image ID
  381. result += "_";
  382. // tmpid_str = std::to_string(outBuff[i][1]);
  383. MS_PRINT("label_classes i %d, outBuff %d",i, (int) outBuff[i][1]);
  384. tmpid_str = label_classes[(int) outBuff[i][1]];
  385. result += tmpid_str; // label id
  386. result += "_";
  387. tmpid_str = std::to_string(outBuff[i][2]);
  388. result += tmpid_str; // scores
  389. result += "_";
  390. tmpid_str = std::to_string(outBuff[i][3]);
  391. result += tmpid_str; // xmin
  392. result += "_";
  393. tmpid_str = std::to_string(outBuff[i][4]);
  394. result += tmpid_str; // ymin
  395. result += "_";
  396. tmpid_str = std::to_string(outBuff[i][5]);
  397. result += tmpid_str; // xmax
  398. result += "_";
  399. tmpid_str = std::to_string(outBuff[i][6]);
  400. result += tmpid_str; // ymax
  401. result += ";";
  402. }
  403. return result;
  404. }
  405. ```