Shulou(Shulou.com)06/01 Report--

What this article shares to you is about how to test the network SSD/Faster R-CNN/YOLO in deep learning. The editor thinks it is very practical, so I share it with you. I hope you can get something after reading this article.

What is testing?

The detection task in deep learning (Detection) is to detect the position of the object in the picture, which generally needs to be framed. For example, in the following picture, people, sheep, and dogs are all framed out. Specifically, the network needs to output the coordinates of the box.

What are the detection networks?

Commonly used detection can be roughly divided into two categories: two stage and one stage. Two stage means that there are two steps in the detection network, the first step is to frame the object, and the second step is to determine the classification of the object. One stage is based on the extracted features and directly uses the regression network to get its classification and detection frame. Among them, Faster R-CNN belongs to two stage,SSD and YOLO belongs to one stage.

The advantage of One stage is that it is fast, and the advantage of Two stage is accuracy. After getting the test box, it is much easier to classify than to directly return to the network to get classification.

The following figure shows the structure of Faster RCNN:

It is divided into four steps:

-convolution layer. The original image first goes through conv-relu-pooling 's multi-layer convolution neural network to extract the feature image (feature map).

-RPN layer, region proposal networks. The RPN layer is used to generate the candidate box, and use softmax to determine whether the candidate box is the foreground or background, select the foreground candidate box (because the object is usually in the foreground), and regress and adjust the position of the candidate box to get the feature map of the object in the frame.

-ROI layer, which converts the feature map pool of objects of different sizes into the same size, and then sends it to the subsequent full connection layer for object classification.

-Classification layer. The feature map output from the ROI layer is used for classification, and the detection box bounding box is regressed again to get a more accurate shape and position.

SSD and YOLO can be compared together, because they both belong to one stage networks, and after obtaining the feature map, detection and classification are carried out together. The difference between SSD and YOLO is that SSD can extract multi-scale features. We can see that the above SSD structure has six arrows pointing to Detections, and each arrow represents a scale of feature map. The more backward the convolution layer, the more you can see the global features. YOLO has no such design, only one feature is output to Detections network.

Because of the design of SSD network, the detection accuracy is higher than YOLO, and the speed is faster than Faster RCNN, so it is also one of the most widely used detection networks. In addition, in the original paper of SSD, VGG skeleton is used to extract features, and now many better feature extraction networks, such as ResNet,mobilenet, are also widely used to further improve accuracy or speed up.

The above is how to detect the network SSD/Faster R-CNN/YOLO in deep learning. The editor believes that there are some knowledge points that we may see or use in our daily work. I hope you can learn more from this article. For more details, please follow the industry information channel.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.