Shulou(Shulou.com)05/31 Report--

The main content of this article is to explain "what is the role of PANet". Interested friends may wish to have a look. The method introduced in this paper is simple, fast and practical. Next, let the editor take you to learn "what is the role of PANet?"

Guide reading

Very simple and efficient feature pyramid module.

Is one of the most important computer vision processes, it divides the image into smaller, multiple segments, so that the target representation and further analysis becomes simple. This process has a variety of applications, from the localization of tumors in medical images and the development of biometric target detection in machine vision. The process of image segmentation is divided into two parts: Semantic segmentation and Instance segmentation.

Semantic segmentation refers to the classification of pixels in an image into meaningful target categories, such as skies, roads or buses.

Case segmentation includes identifying, classifying and locating various instances (objects) in the image at the pixel level, and requires the preservation of the finest features in the image. It is one of the most complex tasks in target detection. In the past, Mask R-CNN was the most commonly used case segmentation technique. The single-stage target detector technology YOLO3 uses the feature pyramid. In a recent version of YOLO, YOLOv4, a new method is used to split instances, called Path Aggregation Network or PANet, or PAN for short. Let's take a closer look at this technology.

PANet:

The PANet bit lies in the neck of the YOLOv4 model, which mainly enhances the instance segmentation process by preserving spatial information.

The properties of PANet

The reason why PANet is selected for instance segmentation in YOLOv4 is that it can accurately save spatial information and help to correctly locate pixels and form mask.

The features that make PANet so accurate are:

1. Path enhancement from bottom to top

When the image passes through each layer of the neural network, the complexity of the feature increases and the spatial resolution of the image decreases. Therefore, pixel-level mask can not be accurately recognized by high-level features.

FPN used in YOLOv3 uses a top-down path to extract semantic-rich features and combine them with accurate location information. But for generating mask for large targets, this approach can cause the path to be too lengthy, because spatial information may need to be propagated to hundreds of layers.

On the other hand, PANet uses another bottom-up path, while FPN uses a top-down path. This helps shorten the path by using a horizontal connection from the bottom to the top. This is called a * * "shortcut" * * connection, and it has only about 10 layers.

two。 Adaptive feature pooling

Previously used techniques, such as Mask-RCNN, use single-stage features to make mask predictions. If the area of interest is large, use ROI Align Pooling to extract features from a higher level. Although quite accurate, this can still lead to unwanted results, because sometimes two proposals are only 10 pixels different, but are assigned to two different layers, when in fact they are very similar proposals.

To avoid this, PANet uses features from all layers and lets the network decide which ones are useful. ROI alignment operation is performed on each feature graph to extract the features of the target. Then there is the maximum fusion operation at the element level to adapt the network to the new features.

3. Fully connected convergence

In Mask-RCNN, FCN is used instead of the full connection layer because it retains spatial information and reduces the number of parameters in the network. However, because the parameters of all spatial locations are shared, the model does not actually learn how to use pixel positions to predict, and by default it displays the sky at the top of the image and the road at the bottom.

On the other hand, the full connection layer is position-sensitive and can adapt to different spatial locations.

PANet uses information from these two layers to provide more accurate mask predictions.

Changes made for YOLOv4

PANet usually uses adaptive feature pooling to add adjacent layers together for mask prediction. However, when using PANet in YOLOv4, this method makes some changes, for example, instead of adding adjacent layers, a splicing operation is applied to them to improve the accuracy of the prediction.

Performance analysis.

Using the ResNet-50 backbone and training with multi-scale images, PANet surpassed the Mask-RCNN and 2016 champions, and also won the 2017 COCO instance segmentation challenge, ranking second in target detection tasks that do not require large batch training.

It has also consistently outperformed Mask-RCNN on Cityscapes datasets. After the pre-training of COCO, the model can be 4.4% higher than Mask-RCNN.

YOLOv4 uses PANet. Because of its simple implementation and high performance, it improves the accuracy of prediction and is twice as fast as EfficientDet.

From the perspective of APs, YOLOv4 achieves an AP value of 43.5% (65.7% AP accuracy) on the MS COCO data set, and achieves a real-time speed of ~ 65 frames per second on Tesla V100, making it the fastest and most accurate detector. The performance is improved by 10-12% due to the inclusion of PANet instead of FPN,YOLOv4 used in YOLOv3!

At this point, I believe that you have a deeper understanding of "what is the role of PANet", you might as well come to the actual operation! Here is the website, more related content can enter the relevant channels to inquire, follow us, continue to learn!

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.