I. Definition
Perceptual field is defined as the area where the input image can be seen according to the characteristics of convolutional neural network. In other words, the output of features is influenced by pixels in the receptive field.
For example, as shown in the figure below (two dimensions are simplified to one dimension for convenience), the convolution kernel of each layer of this three-layer neural convolution neural network is _ = 3. , = 1, the receptive field corresponding to the topmost feature is 7x7 as shown in the figure.
Second, the calculation method
The receptive field of 1 layer [1]
The second layer is characterized by a receptive field of 5.
Second receptive field [1]
The third layer is characterized by a receptive field of 7.
The third receptive field [1]
If there is an extended conv, the calculation formula is
Third, walk up a flight of stairs.
What is described above is the theoretical receptive field, and the effective receptive field (actual effective receptive field) of the feature is actually much smaller than the theoretical receptive field, as shown in the following figure. The specific mathematical analysis is more complicated, so I won't go into details. If you are interested, please refer to paper [2].
Examples of effective receptive fields [2]
Two-layer 3x3 conv calculation flow chart
Fourth, application
classify
Cao Xudong wrote a technical report entitled "Practical Theory of Designing Extremely Deep Convolutional Neural Networks", which mentioned that when designing an image classifier based on deep convolution neural networks, two conditions need to be met to ensure good results:
First of all, for each convolution layer, its ability to learn more complex patterns should be guaranteed; Secondly, the top receptive field should not be larger than the image area.
The second condition is the limitation of the size of the characteristic receptive field of the highest level network of convolutional neural network.
object detection
Nowadays, most popular target detection networks are based on anchor, such as SSD series, yolo after v2 and faster rcnn series.
The target detection network based on anchor points will preset a set of anchor points with different sizes, such as 32x32, 64x64, 128x 128, 256x256. So many anchors, which floor should I put them on? At this time, the size of the site is an important consideration.
The characteristic receptive field of the anchor layer should match the size of the anchor. It is not good to feel that the wild is too big than the anchor, and it is not good to feel too small. If you think the wild one is much smaller than the anchor, it's like giving you only one foot to tell you what kind of bird it is. If you think the wild one is much bigger than the anchor, it's like giving you a map of the world to show you where the Forbidden City is.
The paper "S3FD: One-shot Scale Invariant Face Detector" is an example of designing anchor point size according to receptive field. The original words in the article are as follows
We design the anchor scale according to the effective receptive field.
Face box: the high-precision CPU real-time face detector is based on the same receptive field when designing multi-scale anchors. One contribution of this paper is that
We introduced the multi-scale convolution layer.
(MSCL) deals with various facial proportions through abundance.
Discrete anchors on receptive fields and layers
Quote:
[1] Convolutional neural network
[2] Understand the effective receptive field in deep convolutional neural networks.
Recommended reading:
Pay attention to the official account of SIGAICN WeChat, reply to the article to get the code, and you can get the full-text link.
[1] machine learning-a magnificent 40-year acquisition code SIGAI04 13.
[2] What mathematical knowledge do you need to learn machine learning well? Acquisition code SIGAI04 17.
[3] The evolutionary history acquisition code of face recognition algorithm SIGAI0420.
[4] Overview of target detection algorithms based on deep learning.
[5] Why can convolutional neural networks dominate the field of computer vision? Get the code SIGAI0426.
[6] Understand the context acquisition code SIGAI0428 of SVM with a picture.
[7] Overview of face detection algorithms. Acquisition code SIGAI0503.
[8] Understand the activation function acquisition code SIGAI0505 of neural network.
[9] Evolutionary history and structural improvement of deep convolutional neural network-40 pages comprehensive interpretation of acquisition code SIGAI0508.
[10] Understand the gradient descent method to obtain the code SIGAI05 1 1.
[1 1] An overview of recurrent neural networks-a sharp tool for speech recognition and natural language processing-SIGAI0515.
[12] Understand the convex optimization acquisition code SIGAI05 18.
[13] experimental understanding of SVM kernel function and parameter acquisition code SIGAI0522.
[14] SIGAI summarizes the pedestrian detection algorithm acquisition code SIGAI0525.
[15] Application of machine learning in autonomous driving —— Take Baidu Apollo platform as an example (I) to obtain the code SIGAI0529.
[16] Understanding Newton Method SIGAI 20 18.5.3 1
[17] May Group Topics Highlights-Some Issues Worth Thinking about Machine Learning and Deep Learning Get the code SIGAI060 1.
[18]Adaboost algorithm obtains the code SIGAI0602.
[19] FlowNet to FlowNet2.0: The optical flow prediction algorithm based on convolutional neural network obtained the code SIGAI0604.
[20] Understand the principal component analysis (PCA) acquisition code SIGAI0606.
[2 1] Summary of key points detection of human skeleton. Acquisition code SIGAI0608.
[22] understand the decision tree acquisition code SIGAI06 1 1.
[23] Summarize the commonly used machine learning algorithms in one sentence to obtain the code SIGAI06 13.
[24] Target detection algorithm of YOLO acquisition code SIGAI06 15.
[25] Understand the fitting acquisition code SIGAI06 18.
[26] Understanding Calculation: From √2 to alpha go—— 1 On the acquisition code SIGAI0620 from √2.
[27] scene text detection-introduction-CTPN algorithm obtains code SIGAI0622.
[28] convolutional neural network compression and acceleration acquisition code SIGAI0625.
[29] k nearest neighbor algorithm to obtain the code SIGAI0627.
[30] Overview of natural scene text detection and recognition technology Acquisition code SIGAI0629.
[3 1] Understanding calculation: from √2 to alpha go—— The collection code of historical background of the second season of neural computing SIGAI0702.
[32] Machine learning algorithm graph acquisition code SIGAI0704.
Derivation of Back Propagation Algorithm —— Fully Connected Neural Network Obtaining Code SIGAI0706.
[34] Overview of production countermeasure network model acquisition code SIGAI0709.
[35] How to become an excellent algorithm engineer acquisition code SIGAI07 1 1.
[36] Understanding Calculation: From √2 to AlphaGo-Section III Mathematical Model of Neural Network Obtaining Code SIGAI0702.
[37] Collection code SIGAI6 [37] S3FD face detection algorithm
[38] The crowd counting method based on deep negative correlation learning obtains the code SIGAI07 18.
[39] Overview of manifold learning acquisition code SIGAI0723
Link: /p/7caf6bb6d62 1
Source: