【Android,Kotlin,TFLite】移动设备集成深度学习轻模型TFlite(物体检测篇)

深度学习.Tensorflow.TFLite.移动设备集成深度学习轻模型TFlite.图像分类篇

Why i create it?

​​为了创建一个易用且易于集成的TFlite加载库, 所以TFLiteLoader应运而生​​

• 开源Github项目地址 TFLiteLoader

集成 ObjectDetector

​​依赖​​

allprojects {
  repositories {
    ...
    maven { url 'https://jitpack.io' }
  }
}

dependencies {
  implementation 'com.github.mozhimen.TFLiteLoader:objectdetector:1.0.2'
}

​​接入​​

1. 全局声明

private lateinit var _tfLiteObjectDetector: TFLiteObjectDetector

2. 在onCreate中进行初始化

_tfLiteObjectDetector = TFLiteObjectDetector.create("efficientdet-lite0.tflite", listener = _objectDetectorListener)

3. 异步声明_objectDetectorListener

private val _objectDetectorListener: IObjectDetectorListener = object : IObjectDetectorListener {
    override fun onError(error: String) {
        runOnUiThread {
            error.showToast()
        }
    }

    override fun onResults(imageWidth: Int, imageHeight: Int, inferenceTime: Long, results: MutableList<Detection>?) {
        runOnUiThread {
            results?.let {
                vb.objectDetectionOverlay.setObjectRect(imageWidth, imageHeight, results)
            }
        }
    }
}

3. 物体检测

_tfLiteObjectDetector.detect({你的Bitmap}, 0)

• 对返回数据的处理示例, 可以pull代码参考demo, 这是回调中的处理

override fun onResults(imageWidth: Int, imageHeight: Int, inferenceTime: Long, results: MutableList<Detection>?) {
  runOnUiThread {
      results?.let {
          vb.objectDetectionOverlay.setObjectRect(imageWidth, imageHeight, results)
    }
  }
}

4. 结果

对于返回值的说明

• MutableList{Detection}

@AutoValue
@UsedByReflection("object_detection_jni.cc")
public abstract class Detection {
    public Detection() {
    }

    @UsedByReflection("object_detection_jni.cc")
    public static Detection create(RectF boundingBox, List<Category> categories) {
        return new AutoValue_Detection(new RectF(boundingBox), Collections.unmodifiableList(new ArrayList(categories)));
    }

    //检测物体在画面的位置信息
    public abstract RectF getBoundingBox();

    //类别集合
    public abstract List<Category> getCategories();
}

​​完整demo代码​​

@PermissionKAnnor(permissions = [Manifest.permission.CAMERA])
class ObjectDetectionActivity : BaseKActivity<ActivityObjectDetectionBinding, BaseKViewModel>(R.layout.activity_object_detection) {

    private lateinit var _tfLiteObjectDetector: TFLiteObjectDetector
    private val _objectDetectorListener: IObjectDetectorListener = object : IObjectDetectorListener {
        override fun onError(error: String) {
            runOnUiThread {
                error.showToast()
            }
        }

        override fun onResults(imageWidth: Int, imageHeight: Int, inferenceTime: Long, results: MutableList<Detection>?) {
            runOnUiThread {
                results?.let {
                    vb.objectDetectionOverlay.setObjectRect(imageWidth, imageHeight, results)
                }
            }
        }
    }

    override fun initData(savedInstanceState: Bundle?) {
        PermissionK.initPermissions(this) {
            if (it) {
                initView(savedInstanceState)
            } else {
                PermissionK.applySetting(this)
            }
        }
    }

    override fun initView(savedInstanceState: Bundle?) {
        initLiteLoader()
        initCamera()
    }

    private fun initLiteLoader() {
        _tfLiteObjectDetector = TFLiteObjectDetector.create("efficientdet-lite0.tflite", listener = _objectDetectorListener)
//        _tFLiteLabelImageClassifier = TFLiteLabelImageClassifier.create("?", "labels.txt", modelType = ModelType.QUANTIZED_EFFICIENTNET)
//        _tFImageClassifier = TFImageClassifier.create("output_graph.pb", "output_labels.txt", "input", 299, "output", 128f, 128f, 0.1f, 1)
    }

    private fun initCamera() {
        vb.objectDetectionPreview.initCamera(this, CameraSelector.DEFAULT_BACK_CAMERA)
        vb.objectDetectionPreview.setImageAnalyzer(_frameAnalyzer)
        vb.objectDetectionPreview.startCamera()
    }

    private val _frameAnalyzer: ImageAnalysis.Analyzer by lazy {
        object : ImageAnalysis.Analyzer {
            private val _reentrantLock = ReentrantLock()

            @SuppressLint("UnsafeOptInUsageError", "SetTextI18n")
            override fun analyze(image: ImageProxy) {
                try {
                    _reentrantLock.lock()
                    val bitmap: Bitmap = if (image.format == ImageFormat.YUV_420_888) {
                        ImageConverter.yuv2Bitmap(image)!!
                    } else {
                        ImageConverter.jpeg2Bitmap(image)
                    }
                    val rotateBitmap = UtilKBitmap.rotateBitmap(bitmap, 90)

                    _tfLiteObjectDetector.detect(rotateBitmap, 0)
                } finally {
                    _reentrantLock.unlock()
                }

                image.close()
            }
        }
    }
}

关于这里的框架代码, 可以参考我另一个开源框架库: SwiftKit

• 本示例代码所持引用:

implementation 'com.github.mozhimen.SwiftKit:basick:1.1.1'
implementation('com.github.mozhimen.SwiftKit:abilityk:1.1.1') {
    exclude group: 'com.mozhimen.abilityk.scank'
    exclude group: 'com.huawei.hms'
}
implementation 'com.github.mozhimen.SwiftKit:componentk:1.1.1'

综上所述: 集成是不是很简单, 那赶快试试吧