OpenCV3.3 DNN Introduction

In the field of machine vision, deep learning is now the most popular and fastest growing direction. OpenCV since version 3.1, it has added the DNN module to the contrib. By the 3.3 release, the DNN module was promoted from contrib to the formal code block. The location in the warehouse is: HTTPS://GITHUB.COM/OPENCV/OPENCV/TREE/MASTER/MODULES/DNN. At the same time, compared to the 3.1 version, the 3.3 version of the DNN made a great improvement.

The DNN module, in addition to LIBPROTOBUF, does not rely on any third-party libraries, and LIBPROTOBUF has been included in the OpenCV thirdparty, installation OpenCV will be installed, very convenient.

Currently, the DNN module supports the loading of well-trained models (i.e., these models need to be well-trained in deep learning frameworks such as Caffe, TensorFlow, Torch/pytorch, etc.), and the forward propagation process (i.e. prediction) is performed.

Supported Deep learning libraries:

Caffe 1 TensorFlow Torch/pytorch

the main layers and functions:

Absval (caffe/layers/absval_layer.hpp This layer is relatively simple: the main is to seek absolute value) averagepooling (mean pooling) batchnormalization (like activating function layer, convolutional layer, fully connected layer, The same as the pool layer, the BN also belongs to the network layer, in the network middle layer data to do a normalization processing) concatenation (caffe through the concatenation layer, you can link multiple blobs into a blob) convolution ( including dilated convolution) Crop deconvolution, a.k.a. Transposed convolution or full convolution DetectionOutput (SSD- Specific layer) dropout eltwise (+, *, Max) (Caffe provides an action-by-element level. It supports 3 basic operations: PROD by element, sum by element, Max Save maximum element Flatten (Caffe flattening layer is to put an input size n * c * H * w into a simple vector whose size is n * (c*h*w) * 1 * 1) fullyconnected LRN (LRN in local Response Normalization,caffe is normalized to a local input area) LSTM maxpooling (max pooling) maxunpooling MVN Normalizebbox (ssd-specific layer) Padding permute Power Prelu (including Channelprelu with channel-specific slopes) Prior Box (ssd-specific layer) ReLU RNN scale Shift Sigmoid Slice (Caffe layer's role is to break Slice into multiple bottom as needed) tops (activation function) Softmax T (splitting layer in Caffe can separate an input blob into multiple outputs blobs) TanH (activation function)Some of the networks that have been tested:AlexNet Googlenet V1 (also referred to as inception-5h) resnet-34/50/... Squeezenet v1.1 vgg-based FCN (semantical segmentation Network) ENet (lightweight semantical Segmentation network) Vgg-bas Ed SSD (Object detection Network) mobilenet-based SSD (Light-weight object Detection NetworkRoutines :Https://github.com/opencv/opencv/tree/master/samples/dnn