Tensorflow 中用于 CNN 的 Gabor 滤波器，python答案

【问题标题】：Gabor Filters for CNN in Tensorflow, pythonTensorflow 中用于 CNN 的 Gabor 滤波器，python
【发布时间】：2020-05-27 10:05:52
【问题描述】：

我想在 CNN 中使用 Gabor 过滤器作为内核，但我找不到解决方案。我找到但不起作用的东西。

import tensorflow as tf
import cv2
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D, BatchNormalization
from tensorflow.keras.layers import Activation, Flatten, Dropout, Conv2DTranspose, LeakyReLU, Concatenate, Lambda

from tensorflow.keras import backend as K

def get_gabor_tensor(ksize, sigmas, thetas, lambdas, gammas, psis):
    n_kernels = len(sigmas) * len(thetas) * len(lambdas) * len(gammas) * len(psis)
    gabors = []
    for sigma in sigmas:
        for theta in thetas:
            for lambd in lambdas:
                for gamma in gammas:
                    for psi in psis:
                        params = {'ksize': ksize, 'sigma': sigma,
                                  'theta': theta, 'lambd': lambd,
                                  'gamma': gamma, 'psi': psi}
                        gf = cv2.getGaborKernel(**params, ktype=cv2.CV_32F)
                        gf = K.expand_dims(gf, -1)
                        gabors.append(gf)
    assert len(gabors) == n_kernels
    print(f"Created {n_kernels} kernels.")
    return K.stack(gabors, axis=-1)


def convolve_tensor(x, kernel_tensor=None):
    return K.conv2d(x, kernel_tensor, padding='same')


def gabor_layer(layer, n_filters=16, kernel_size=3):
    ksize=(3, 3)
    sigmas = [1, 2, 3, 4]
    thetas = np.linspace(0, np.pi, 4, endpoint=False)
    lambdas=[8, 16, 32, 64]
    psis = np.linspace(0, 2*np.pi, 2, endpoint=False)
    gammas = np.linspace(1, 0, 2, endpoint=False)

    tensor = get_gabor_tensor(ksize, sigmas, thetas, lambdas, gammas, psis)

    x = Lambda(convolve_tensor, arguments={'kernel_tensor': tensor})(layer)
    c1 = Conv2D(filters=16, kernel_size=(3, 3), padding='same')(layer)
    p1 = MaxPooling2D((2, 2))(c1)
    output = Dropout(0.1)(p1)
    return output

还有错误

ValueError: Depth of output (256) is not a multiple of the number of groups (3) for 'lambda/convolution' (op: 'Conv2D') with input shapes: [?,96,96,3], [3,3,1,256].

解决方案来自https://github.com/bdevans/GaborNet/blob/master/gabornet.py

【问题讨论】：

标签： python tensorflow keras deep-learning gabor-filter

【解决方案1】：

通过 Input(shape=(128,192,1)) ;即带有灰色图像。我遇到了同样的错误，它得到了解决。

【讨论】：