卷积网络 感受野计算

最新推荐文章于 2022-08-30 18:26:16 发布
最新推荐文章于 2022-08-30 18:26:16 发布
阅读量392 收藏
点赞数 1
文章标签： python 感受野 神经网络
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/weixin_43951539/article/details/100759414
版权

理论感受野计算

在设计网络层次时，有效感受野是一个重要参考依据，但是很多资料并没有把感受野的计算讲清楚，所以正好自己总结一下，并给出分类网络与目标检测网络的理论感受野计算的代码。
有效感受野比理论感受野小。

特征的有效感受野（实际起作用的感受野）实际上是小于理论感受野的.虽然x11和x33都位于第二层特征感受野内，但是二者对最后的特征o11(2)的影响却不同，输入中越靠感受野中间的元素对特征的贡献越大。

具体可参见论文Understanding the Effective Receptive Field

计算理论感受野，分为2步

  1. 依次列出每个卷积层的[filter_size 卷积核大小, stride 步长], 如果同时想计算每层的大小需要补充 [padding 填充]，如class networkReceptiveField 中我分别列出了Alexnet， vgg16， 基于darknet53的yolo v3，基于mobilenet v2的yolo v3的卷积层结构[filter_size, stride, padding]在self.networks

  2. 倒推计算卷积层感受野，从网络的最后一层逐层递推到该层。如下图，计算第三层实心点的理论感受野，先计算该点对应第二层的感受野，再计算对应第一层的感受野。（1，2，3层特征图大小6*6，padding为1，stride为1）

receptive_field = 1
for layer in reversed(range(layernum)):
    filter_size, stride, _ = net[layer]
    receptive_field = ((receptive_field - 1) * stride) + filter_size

补充：卷积层padding的计算

  • 卷积中padding的计算公式
    p a d d i n g = ( s h a p e o u t − 1 ) ∗ s t r i d e + f i l t e r s i z e − s h a p e i n 2 padding = \frac{(shape_{out} - 1) * stride + filter_{size} - shape_{in}}{2} padding=2(shapeout​−1)∗stride+filtersize​−shapein​​
  • 如果特征图大小不变
    p a d d i n g = f i l t e r s i z e − 1 2 padding = \frac{filter_{size} - 1}{2} padding=2filtersize​−1​

完整的代码如下

import numpy as np


class networkReceptiveField(object):
    def __init__(self, img_input_size):
        # by default, provide 4 networks structures
        self.networks = {
        # parameters: [filter_size, stride, padding]
    'alexnet': {'parameters': [[11, 4, 0], [3, 2, 0], [5, 1, 2], [3, 2, 0], [3, 1, 1], [3, 1, 1], [3, 1, 1], [3, 2, 0]],
                'name': ['conv1', 'pool1', 'conv2', 'pool2', 'conv3', 'conv4', 'conv5', 'pool5']},

    'vgg16':   {'parameters': [[3, 1, 1], [3, 1, 1], [2, 2, 0], [3, 1, 1], [3, 1, 1], [2, 2, 0], [3, 1, 1], [3, 1, 1], [3, 1, 1],
                      [2, 2, 0], [3, 1, 1], [3, 1, 1], [3, 1, 1], [2, 2, 0], [3, 1, 1], [3, 1, 1], [3, 1, 1],
                      [2, 2, 0]],
                'name': ['conv1_1', 'conv1_2', 'pool1', 'conv2_1', 'conv2_2', 'pool2', 'conv3_1', 'conv3_2',
                         'conv3_3', 'pool3', 'conv4_1', 'conv4_2', 'conv4_3', 'pool4', 'conv5_1', 'conv5_2', 'conv5_3',
                         'pool5']},

    'darknet53': {'parameters': [
                           # stage 0
                          [3, 1, 1], [3, 2, 0.5],
                           # stage 1
                          [1, 1, 0], [3, 1, 1], [3, 2, 0.5],
                           # stage 2
                          [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [3, 2, 0.5],
                           # stage 3
                          [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1],
                          [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1],
                          # [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], # yolo layers S
                           # stage 4
                          [3, 2, 0.5], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0],
                          [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1],
                          # [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], # yolo layers M
                           # stage 5
                          [3, 2, 0.5], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1],
                          [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0] # yolo layers L
                          ],
                  'name': ['conv00', 'conv00d',
                           'conv10', 'conv11', 'conv11d',
                           'conv20', 'conv21', 'conv22', 'conv23', 'conv23d',
                           'conv30', 'conv31', 'conv32', 'conv33', 'conv34', 'conv35', 'conv36', 'conv37',
                           'conv38', 'conv39', 'conv310', 'conv311', 'conv312', 'conv313', 'conv314', 'conv315',
                           # 'yoloS0', 'yoloS1', 'yoloS2', 'yoloS3', 'yoloS4', 'yoloS5', 'yoloS6',
                           'conv40d', 'conv40', 'conv41', 'conv42', 'conv43', 'conv44', 'conv45', 'conv46',
                           'conv47', 'conv48', 'conv49', 'conv410', 'conv411', 'conv412', 'conv413', 'conv414', 'conv415',
                           # 'yoloM0', 'yoloM1', 'yoloM2', 'yoloM3', 'yoloM4', 'yoloM5', 'yoloM6',
                           'conv50d', 'conv50', 'conv51', 'conv52', 'conv53', 'conv50', 'conv51', 'conv52', 'conv53',
                           'yoloL0', 'yoloL1', 'yoloL2', 'yoloL3', 'yoloL4', 'yoloL5', 'yoloL6']},
                           
    'mobilenetv2': {'parameters': [
                        # stage 0
                        [3, 2, 0.5],
                        # stage 1
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 2, 0.5], [1, 1, 0],
                        # stage 2
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 2, 0.5], [1, 1, 0],
                        # stage 3
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0],
                        # [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], # yolo layers S
                        [1, 1, 0], [3, 2, 0.5], [1, 1, 0],
                        # stage 4
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0],
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0],
                        # [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0],  # yolo layers M
                        [1, 1, 0], [3, 2, 0.5], [1, 1, 0],
                        # stage 5
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0], [3, 1, 1], [1, 1, 0], [1, 1, 0],
                        [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0], [3, 1, 1], [1, 1, 0]  # yolo layers L
    ],
        'name': ['conv00d',
                 'conv10', 'conv11', 'conv12', 'conv13', 'conv14', 'conv15d',
                 'conv20', 'conv21', 'conv22', 'conv23', 'conv24', 'conv25d',
                 'conv30', 'conv31', 'conv32', 'conv33', 'conv34', 'conv35',
                 # 'yoloS0', 'yoloS1', 'yoloS2', 'yoloS3', 'yoloS4', 'yoloS5', 'yoloS6',
                 'conv36', 'conv37', 'conv38d',
                 'conv40', 'conv41', 'conv42', 'conv43', 'conv44', 'conv45', 'conv46', 'conv47', 'conv48',
                 'conv49', 'conv410', 'conv411', 'conv412', 'conv413', 'conv414', 'conv415', 'conv416', 'conv417',
                 # 'yoloM0', 'yoloM1', 'yoloM2', 'yoloM3', 'yoloM4', 'yoloM5', 'yoloM6',
                 'conv418', 'conv419', 'conv420d',
                 'conv50', 'conv51', 'conv52', 'conv53', 'conv54', 'conv55', 'conv56', 'conv57', 'conv58', 'conv59',
                 # 'yoloL0', 'yoloL1', 'yoloL2', 'yoloL3', 'yoloL4', 'yoloL5', 'yoloL6'
                 ]}}

        self.imsize = img_input_size

    def compute_outsize(self, net, layernum):
        insize = np.array(self.imsize, dtype=int)
        outsize = []
        stride_total = 1
        for layer in range(layernum):
            filter_size, stride, padding = net[layer]
            outsize = (insize - filter_size + 2 * padding) / stride + 1
            insize = outsize
            stride_total = stride_total * stride
        return outsize, stride_total

    def inFromOut(self, net, layernum):
        receptive_field = 1
        for layer in reversed(range(layernum)):
            filter_size, stride, padding = net[layer]
            receptive_field = ((receptive_field - 1) * stride) + filter_size
        return receptive_field

    def get_receptive_field(self, net):
        print("neutral network feeded with image size: {} * {}".format(self.imsize[0], self.imsize[1]))
        # choose one network to analyse it's respective field
        network = self.networks[net]['parameters']
        print(net)
        for i in range(len(network)):
            outsize, stride_total = self.compute_outsize(network, i + 1)
            receptive_field = self.inFromOut(network, i + 1)
            print("Layer Name = %8s, Output size = %4.1d * %3.1d, Stride = %2.1d, RF size = %2.1d" % (
                self.networks[net]['name'][i], outsize[0], outsize[1], stride_total, receptive_field))

if __name__ == '__main__':
    imsize = [832, 1024]
    networks = networkReceptiveField(imsize)
    # alexnet, vgg16, darknet53(with yolov3 layers), mobilenetv2(with yolov3 layers)
    # in order to analyse the full receptive field with yolo layer,
    # you need uncomment the corresponding layers in self.networks 'parameters'
    networks.get_receptive_field('vgg16')

layer output sizes given image = 1024 x 1024
****************** YOLO V3 with Backbone Darknet53*******************
stage 0 ============================================================
Layer Name =   conv00, Output size = 1024, Stride =  1, RF size =  3
Layer Name =  conv00d, Output size =  512, Stride =  2, RF size =  5
stage 1 ============================================================
Layer Name =   conv10, Output size =  512, Stride =  2, RF size =  5
Layer Name =   conv11, Output size =  512, Stride =  2, RF size =  9
Layer Name =  conv11d, Output size =  256, Stride =  4, RF size = 13
stage 2 ============================================================
Layer Name =   conv20, Output size =  256, Stride =  4, RF size = 13
Layer Name =   conv21, Output size =  256, Stride =  4, RF size = 21
Layer Name =   conv22, Output size =  256, Stride =  4, RF size = 21
Layer Name =   conv23, Output size =  256, Stride =  4, RF size = 29
Layer Name =  conv23d, Output size =  128, Stride =  8, RF size = 37
stage 3 ============================================================
Layer Name =   conv30, Output size =  128, Stride =  8, RF size = 37
Layer Name =   conv31, Output size =  128, Stride =  8, RF size = 53
Layer Name =   conv32, Output size =  128, Stride =  8, RF size = 53
Layer Name =   conv33, Output size =  128, Stride =  8, RF size = 69
Layer Name =   conv34, Output size =  128, Stride =  8, RF size = 69
Layer Name =   conv35, Output size =  128, Stride =  8, RF size = 85
Layer Name =   conv36, Output size =  128, Stride =  8, RF size = 85
Layer Name =   conv37, Output size =  128, Stride =  8, RF size = 101
Layer Name =   conv38, Output size =  128, Stride =  8, RF size = 101
Layer Name =   conv39, Output size =  128, Stride =  8, RF size = 117
Layer Name =  conv310, Output size =  128, Stride =  8, RF size = 117
Layer Name =  conv311, Output size =  128, Stride =  8, RF size = 133
Layer Name =  conv312, Output size =  128, Stride =  8, RF size = 133
Layer Name =  conv313, Output size =  128, Stride =  8, RF size = 149
Layer Name =  conv314, Output size =  128, Stride =  8, RF size = 149
Layer Name =  conv315, Output size =  128, Stride =  8, RF size = 165
*****************************branch 1********************************
Layer Name =   yoloS0, Output size =  128, Stride =  8, RF size = 165
Layer Name =   yoloS1, Output size =  128, Stride =  8, RF size = 181
Layer Name =   yoloS2, Output size =  128, Stride =  8, RF size = 181
Layer Name =   yoloS3, Output size =  128, Stride =  8, RF size = 197
Layer Name =   yoloS4, Output size =  128, Stride =  8, RF size = 197
Layer Name =   yoloS5, Output size =  128, Stride =  8, RF size = 213
Layer Name =   yoloS6, Output size =  128, Stride =  8, RF size = 213
stage 4 ============================================================
Layer Name =  conv40d, Output size =   64, Stride = 16, RF size = 181
Layer Name =   conv40, Output size =   64, Stride = 16, RF size = 181
Layer Name =   conv41, Output size =   64, Stride = 16, RF size = 213
Layer Name =   conv42, Output size =   64, Stride = 16, RF size = 213
Layer Name =   conv43, Output size =   64, Stride = 16, RF size = 245
Layer Name =   conv44, Output size =   64, Stride = 16, RF size = 245
Layer Name =   conv45, Output size =   64, Stride = 16, RF size = 277
Layer Name =   conv46, Output size =   64, Stride = 16, RF size = 277
Layer Name =   conv47, Output size =   64, Stride = 16, RF size = 309
Layer Name =   conv48, Output size =   64, Stride = 16, RF size = 309
Layer Name =   conv49, Output size =   64, Stride = 16, RF size = 341
Layer Name =  conv410, Output size =   64, Stride = 16, RF size = 341
Layer Name =  conv411, Output size =   64, Stride = 16, RF size = 373
Layer Name =  conv412, Output size =   64, Stride = 16, RF size = 373
Layer Name =  conv413, Output size =   64, Stride = 16, RF size = 405
Layer Name =  conv414, Output size =   64, Stride = 16, RF size = 405
Layer Name =  conv415, Output size =   64, Stride = 16, RF size = 437
*****************************branch 2********************************
Layer Name =   yoloM0, Output size =   64, Stride = 16, RF size = 437
Layer Name =   yoloM1, Output size =   64, Stride = 16, RF size = 469
Layer Name =   yoloM2, Output size =   64, Stride = 16, RF size = 469
Layer Name =   yoloM3, Output size =   64, Stride = 16, RF size = 501
Layer Name =   yoloM4, Output size =   64, Stride = 16, RF size = 501
Layer Name =   yoloM5, Output size =   64, Stride = 16, RF size = 533
Layer Name =   yoloM6, Output size =   64, Stride = 16, RF size = 533
stage 5 ============================================================
Layer Name =  conv50d, Output size =   32, Stride = 32, RF size = 469
Layer Name =   conv50, Output size =   32, Stride = 32, RF size = 469
Layer Name =   conv51, Output size =   32, Stride = 32, RF size = 533
Layer Name =   conv52, Output size =   32, Stride = 32, RF size = 533
Layer Name =   conv53, Output size =   32, Stride = 32, RF size = 597
Layer Name =   conv50, Output size =   32, Stride = 32, RF size = 597
Layer Name =   conv51, Output size =   32, Stride = 32, RF size = 661
Layer Name =   conv52, Output size =   32, Stride = 32, RF size = 661
Layer Name =   conv53, Output size =   32, Stride = 32, RF size = 725
*****************************branch 3********************************
Layer Name =   yoloL0, Output size =   32, Stride = 32, RF size = 725
Layer Name =   yoloL1, Output size =   32, Stride = 32, RF size = 789
Layer Name =   yoloL2, Output size =   32, Stride = 32, RF size = 789
Layer Name =   yoloL3, Output size =   32, Stride = 32, RF size = 853
Layer Name =   yoloL4, Output size =   32, Stride = 32, RF size = 853
Layer Name =   yoloL5, Output size =   32, Stride = 32, RF size = 917
Layer Name =   yoloL6, Output size =   32, Stride = 32, RF size = 917

layer output sizes given image = 1024 x 1024
************ YOLO V3 with backbone mobilenetv2 **************
stage 0 ============================================================
Layer Name =  conv00d, Output size =  512, Stride =  2, RF size =  3
stage 1 ============================================================
Layer Name =   conv10, Output size =  512, Stride =  2, RF size =  3
Layer Name =   conv11, Output size =  512, Stride =  2, RF size =  7
Layer Name =   conv12, Output size =  512, Stride =  2, RF size =  7
Layer Name =   conv13, Output size =  512, Stride =  2, RF size =  7
Layer Name =   conv14, Output size =  256, Stride =  4, RF size = 11
Layer Name =  conv15d, Output size =  256, Stride =  4, RF size = 11
stage 2 ============================================================
Layer Name =   conv20, Output size =  256, Stride =  4, RF size = 11
Layer Name =   conv21, Output size =  256, Stride =  4, RF size = 19
Layer Name =   conv22, Output size =  256, Stride =  4, RF size = 19
Layer Name =   conv23, Output size =  256, Stride =  4, RF size = 19
Layer Name =   conv24, Output size =  128, Stride =  8, RF size = 27
Layer Name =  conv25d, Output size =  128, Stride =  8, RF size = 27
stage 3 ============================================================
Layer Name =   conv30, Output size =  128, Stride =  8, RF size = 27
Layer Name =   conv31, Output size =  128, Stride =  8, RF size = 43
Layer Name =   conv32, Output size =  128, Stride =  8, RF size = 43
Layer Name =   conv33, Output size =  128, Stride =  8, RF size = 43
Layer Name =   conv34, Output size =  128, Stride =  8, RF size = 59
Layer Name =   conv35, Output size =  128, Stride =  8, RF size = 59
*****************************branch 1********************************
Layer Name =   yoloS0, Output size =  128, Stride =  8, RF size = 59
Layer Name =   yoloS1, Output size =  128, Stride =  8, RF size = 75
Layer Name =   yoloS2, Output size =  128, Stride =  8, RF size = 75
Layer Name =   yoloS3, Output size =  128, Stride =  8, RF size = 91
Layer Name =   yoloS4, Output size =  128, Stride =  8, RF size = 91
Layer Name =   yoloS5, Output size =  128, Stride =  8, RF size = 107
Layer Name =   yoloS6, Output size =  128, Stride =  8, RF size = 107

Layer Name =   conv36, Output size =  128, Stride =  8, RF size = 59
Layer Name =   conv37, Output size =   64, Stride = 16, RF size = 75
Layer Name =  conv38d, Output size =   64, Stride = 16, RF size = 75
stage 4 ============================================================
Layer Name =   conv40, Output size =   64, Stride = 16, RF size = 75
Layer Name =   conv41, Output size =   64, Stride = 16, RF size = 107
Layer Name =   conv42, Output size =   64, Stride = 16, RF size = 107
Layer Name =   conv43, Output size =   64, Stride = 16, RF size = 107
Layer Name =   conv44, Output size =   64, Stride = 16, RF size = 139
Layer Name =   conv45, Output size =   64, Stride = 16, RF size = 139
Layer Name =   conv46, Output size =   64, Stride = 16, RF size = 139
Layer Name =   conv47, Output size =   64, Stride = 16, RF size = 171
Layer Name =   conv48, Output size =   64, Stride = 16, RF size = 171
Layer Name =   conv49, Output size =   64, Stride = 16, RF size = 171
Layer Name =  conv410, Output size =   64, Stride = 16, RF size = 203
Layer Name =  conv411, Output size =   64, Stride = 16, RF size = 203
Layer Name =  conv412, Output size =   64, Stride = 16, RF size = 203
Layer Name =  conv413, Output size =   64, Stride = 16, RF size = 235
Layer Name =  conv414, Output size =   64, Stride = 16, RF size = 235
Layer Name =  conv415, Output size =   64, Stride = 16, RF size = 235
Layer Name =  conv416, Output size =   64, Stride = 16, RF size = 267
Layer Name =  conv417, Output size =   64, Stride = 16, RF size = 267
*****************************branch 2********************************
Layer Name =   yoloM0, Output size =   64, Stride = 16, RF size = 267
Layer Name =   yoloM1, Output size =   64, Stride = 16, RF size = 299
Layer Name =   yoloM2, Output size =   64, Stride = 16, RF size = 299
Layer Name =   yoloM3, Output size =   64, Stride = 16, RF size = 331
Layer Name =   yoloM4, Output size =   64, Stride = 16, RF size = 331
Layer Name =   yoloM5, Output size =   64, Stride = 16, RF size = 363
Layer Name =   yoloM6, Output size =   64, Stride = 16, RF size = 363

Layer Name =  conv418, Output size =   64, Stride = 16, RF size = 267
Layer Name =  conv419, Output size =   32, Stride = 32, RF size = 299
Layer Name = conv420d, Output size =   32, Stride = 32, RF size = 299
stage 5 ============================================================
Layer Name =   conv50, Output size =   32, Stride = 32, RF size = 299
Layer Name =   conv51, Output size =   32, Stride = 32, RF size = 363
Layer Name =   conv52, Output size =   32, Stride = 32, RF size = 363
Layer Name =   conv53, Output size =   32, Stride = 32, RF size = 363
Layer Name =   conv54, Output size =   32, Stride = 32, RF size = 427
Layer Name =   conv55, Output size =   32, Stride = 32, RF size = 427
Layer Name =   conv56, Output size =   32, Stride = 32, RF size = 427
Layer Name =   conv57, Output size =   32, Stride = 32, RF size = 491
Layer Name =   conv58, Output size =   32, Stride = 32, RF size = 491
Layer Name =   conv59, Output size =   32, Stride = 32, RF size = 491
*****************************branch 3********************************
Layer Name =   yoloL0, Output size =   32, Stride = 32, RF size = 491
Layer Name =   yoloL1, Output size =   32, Stride = 32, RF size = 555
Layer Name =   yoloL2, Output size =   32, Stride = 32, RF size = 555
Layer Name =   yoloL3, Output size =   32, Stride = 32, RF size = 619
Layer Name =   yoloL4, Output size =   32, Stride = 32, RF size = 619
Layer Name =   yoloL5, Output size =   32, Stride = 32, RF size = 683
Layer Name =   yoloL6, Output size =   32, Stride = 32, RF size = 683
hoheC+
关注
  • 1
    点赞
  • 踩
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
量化噪声：公式SNR = 6.02 N + 1.76 dB 的扩展推导
09-14
量化噪声：公式SNR = 6.02 N + 1.76 dB 的扩展推导
Dilated Convolution和普通感受野计算的python实现
01-06
这里对Dilated/Atrous Convolution的感受野计算也进行下实现 其实Dilated/Atrous Convolution理解起来很简单多孔卷积，仅仅是为了增加感受野，参数不变 当dilation = 1时就和普通卷积一样 如果使用了dilation卷积...
dilated convolution (扩张卷积) 输入输出特征图尺寸大小关系以及 receptive field (感受野) 计算
xiaojiajia007的博客
06-26 1万+
1. 输入输入出特征图尺寸大小的关系 更多例子可以参考：https://zhuanlan.zhihu.com/p/56940729 带有扩张卷积的输入与输出特征图feature map的大小关系为： 其中，p->padding, d->dilation, k->kernel size，如果是常规的卷积，则另d=1即可。上面这个公式非常得有用，需要牢记！这个关系式推导非常...
『拨云见日』卷积网络感受野的计算
qq_39740996的博客
12-27 493
写在最前，正向感受野计算公式rn=rn−1+(k−1)∗∏i=1n−1sir_n=r_{n-1}+(k-1)*\quad \prod_{i=1}^{n-1} {s_i} \quadrn​=rn−1​+(k−1)∗i=1∏n−1​si​其中k为卷积核尺寸，s为步长。几个博客把这个公式抄来抄去，也没个所以然，实在是头疼看不懂，所以推导一下。 先简单介绍一下普通卷积和扩张卷积： 普通卷积： 卷积后图像的...
深度学习中卷积后的feature map和感受野（receptive fields）大小计算
落日之城
09-21 722
文章目录卷积后Feature map大小的计算感受野感受野是什么？感受野的计算影响感受野大小的因素 卷积后Feature map大小的计算 假设上一层feature map的大小是FnF_{n}Fn​, 那么我们可以用下面的公式来计算下一层卷积后feature map的大小Fn+1F_{n+1}Fn+1​： Fn+1=Fn−ksizestride+1F_{n+1}=\frac{F_{n}-ksiz...
你知道如何计算CNN感受野吗？这里有一份详细指南
小白将
06-11 5754
码字不易，欢迎给个赞！ 欢迎交流与转载，文章会同步发布在公众号：机器学习算法全栈工程师(Jeemy110) 固定大小的CNN特征图可视化 感受野计算公式 附文 小结 本文翻译自A guide to receptive field arithmetic for Convolutional Neural Networks，原作者保留版权。 感受野（receptive field，...
感受野的计算与理解（一看就会）
01-06
A1：（1）从局部理解：（上层->本层）感受野等价与卷积核尺寸。  （2）从全局理解：指输出特征层中一个点映射在原始输入图像中的区域。  Tips：通常感受野指全局理解。 Q2：如何计算一个点映射在原始图像中的区域...
卷积神经网络知识学习笔记
02-24
本文来自简书，本文主要讲解了卷积神经网络知识，包括卷积层和池化层的介绍，希望对您的学习有所帮助。卷积神经网络（CNN）是含有卷...影响元素的前向计算的所有可能输入区域（可能大于输入的实际尺寸）叫做的感受野
感受野特征增强的SSD目标检测算法
05-06
该算法将不同层及不同尺度的特征图以轻量级的方式相融合，下采样层生成新的特征图，通过引入感受野模块，提高网络的特征提取能力，增强特征的表征能力和鲁棒性.和传统SSD算法相比，本文算法在精度上有明显提升，同时...
卷积神经网络（CNN）概念理解
10-26
卷积神经网络基础概念理解： 1.图片知识：通道、位深 2.卷积、卷积核、感受野、梯度、激活函数 3.图片通道提取、图片卷积计算
图像分割--Dilated Convolutions（扩张卷积）
zk_ken的博客
05-25 4224
参考资料：https://blog.csdn.net/zizi7/article/details/77369945                  https://blog.csdn.net/Quincuntial/article/details/78743033以FCN和SegNet 为代表的图像分割方法与传统的 CNN 一样，先对图像做卷积再做池化。其中池化的意义一方面是降低计算量，另一方面...
卷积神经网络里面的感受野计算
slamer111的博客
08-30 2815
卷积神经网络里面的感受野计算
卷积神经网络感受野的计算
baolinq的博客
05-22 3543
神经网络感受野计算 普通卷积感受野计算 感受野指的是当前的特征图的一个特征点在输入空间影响的区域，该点的值被输入空间的这个区域影响，与其他区域无5173。通常有两种方式可以计算感受野，从后往前算和从前往后算。从前往后计算的方法很简单，基本上一个公式就可以计算，使用递推公式计算。 L表示感受野的大小，k表示网络的层序号，f表示卷积核尺寸，s表示步长大小。递推公式的增量为后一部分，对于第...
如何理解扩张卷积(dilated convolution)
热门推荐
zhangxu
08-16 4万+
扩张卷积（Dilated Convolution）也被称为空洞卷积或者膨胀卷积，是在标准的卷积核中注入空洞，以此来增加模型的感受野（reception field）。相比原来的正常卷积操作，扩张卷积多了一个参数： dilation rate，指的是卷积核的点的间隔数量，比如常规的卷积操作dilatation rate为1。它的论文是：《MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS》
卷积神经网络感受野计算指南
小白学视觉
07-23 459
点击上方“小白学视觉”，选择加"星标"或“置顶” 重磅干货，第一时间送达作者：Dang Ha The Hien编译：ronghuaiyang导读之前分享过一篇文章，是关于感受野的，今天这篇文章是专门讲解如何计算感受野的，可以和之前那篇结合起来看，之前那篇文章我也再推送一次给大家。感受野可能是卷积神经网络(CNNs)中最重要的概念之一，值得更多的关注。所有最先进的物体识别方...
【深度学习】感受野
筱
03-21 4340
感受野是卷积神经网络每一层输出的特征图（Feature Map）上的像素点在原始图像上映射的区域大小。 卷积神经网络由输入维度计算输出维度： output size = ( input size - kernel size + 2 * padding ) / stride + 1 卷积神经网络由输出维度计算输入维度： input size = （output size - 1）* strid...
量化噪声(Quantization Noise)
qq_38453556的博客
09-29 1万+
参考资料：http://www.onmyphd.com/?p=quantization.noise.snr 在ADC当中，众多模拟量会被转化成同一种数字量，而对应同一个数字量的模拟量之间的最大差值被称为Q，并且Q=LSB。模拟输入和数字输出之间的差值被称为量化误差（quantization error），量化误差的范围是±Q/2。 假设量化误差在±Q/2的范围内平均分布，...
卷积网络：关于感受野的计算以及卷积网络涉及到的计算
hands_up_down的博客
12-12 687
https://www.bilibili.com/video/BV1q7411T7Y6?from=search&seid=5647822811307033095&spm_id_from=333.337.0.0https://www.bilibili.com/video/BV1q7411T7Y6?from=search&seid=5647822811307033095&spm_id_from=333.337.0.0仔细想想，也许我在之前的文章中对于感受野的理解是错的：认为fea
卷积神经网络中的感受野
最新发布
03-14
在卷积神经网络（Convolutional Neural Network，CNN）中，感受野（Receptive Field）是指在输入图像上，一个特定的神经元对应的输入区域大小。感受野的大小决定了神经元对输入图像的局部信息感知能力。 在CNN中，...

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值