英语写作错误笔记一

of EFP, of the ACT :

of using the EFP, of using ACT;

The STFT of a signal: 

The STFT;

added the Table. I of the revised manuscript: 

added Table. I in the revised manuscript

the convolutional layers: 

convolutional layers

number of overlapped samples:

the number of overlapped samples

To achieve our goal of the separation of anisotropic Gaussian window

To achieve our goal of separation of the anisotropic Gaussian window

revised the Fig. 10

revised Fig. 10

RBF kernal

the RBF kernal

in the each image

in each image

achieves the higher classification accuracy than

achieves higher classification accuracy than

of the specific stage

of a specific stage

MobileNet extracts features from the named "conv2d 11" layer whose the size of output feature maps

MobileNet extracts features from the "conv2d 11" layer whose size of output feature maps

but number of rows is given:

but the number of rows is defined as follows. 

An architecture of TFFNet.

when you use "An", it means that it is not necessarily architecture that you are using.

The architecture of TFFNet.

use the same time range to [0,0.75] s : 

use the same time range [0,0.75] s

The observed signal has a length of 150000 samples in 0.75 seconds

The observed signal has a length of 150000 samples within  a time interval of 0.75 seconds

traditional FPN of construction a multi-resolution

traditional FPN for construction of a multi-resolution

classier' mAP on the UWA communication signals dataset :

classier' mAP for the UWA communication signals dataset 

we revised the manuscript the following text:

we revised the manuscript as follows

We have revised the text as follows.

We have revised the following text.

We have revised classifier to classifier.

This has been corrected. 

The classication performance of the TFFNet is compared with two machine learning methods, random forest (RF)

The classication performance of the TFFNet is compared with that of  two machine learning methods, random forest (RF)

image with 299 * 299 input size is

image of a size 299 *299  is

is lower than RF and SVM-RBF

is lower than that of RF and SVM-RBF

The RBF classification is less accurate compared SVM-RBF.

The RBF classification is less accurate compared to SVM-RBF.

In the work/ Using the way

In this work/ Using this way

get the data

extract the data

TFFNet with STFT get a lower mAP

TFFNet with STFT results in  a lower mAP

references [21][22][23] of the revised manuscript

references [21][22][23] in the revised manuscript

traditional FPN of construction a multi-resolution

traditional FPN for construction of a multi-resolution

section II-A and section II-B

section II

sparse ACT is exible in adjustment between complexity and energy concentration

allows adjustment of the trade-off between complexity and energy concentration

a Beluga whale and a sperm whale sounds

Beluga whale and sperm whale sounds

a list of, values with 

is equal to in the range from 0.2 to 2.6 with an interval of 0.3

Parameters and each is a list of values [0.2, 0.5, 0.8, 1.1, 1.4, 1.7, 2.0,2.3].

The factor  don't contain 

The factor  does not contain 

Consider the high efficiency of sparse ACT

For the high efficiency of sparse ACT

should be big values

should be large values

second one more impulses (spermwhales's clicks)

second one with impulses (spermwhales's clicks)

Using the same hardware configuration, we did not compare with FSST

We could not do this for  FSST

experiment can't be executed

experiment cannot be executed

a key factor in influence of the classification

a key factor influencing the classification

we change the coordinate system [31].

we change the coordinate system as in  [31]

are learnt from some UWA signals.

are learnt from UWA signals.

[12] has ...

The work [12] has ...

An anisotropic operator is selected. 

The parameter is selected  ( is a parameter, not an operator)

the total samples is 1797

total number of samples 

sample number of each class is about 180.

the number of samples within each class

RBF kernel in an example where the numberThere are several manners in

There are several ways in

The drawback of these methods is that the addition of more training data

The drawback of these methods is their low capacity so that  the addition of more training data

a parameter you can tune.

a parameter one can tune.

MobileNet, VGG-16 and Inception V3 extract features

MobileNet, VGG-16 and Inception V3 backbone networks  extract features

The number of samples of ve modulation types is equivalent, 500 training, 110 validation, and 70 test sets.

There are 680 signals with each type of modulation, of which 500, 110, and 70 are used for training, validation and testing, respectively.  

Comparison performance for di erent classification tasks.

Performance comparison for di erent classification tasks.

three trained classiers of RF, SVM-RBF and TFFNet

three trained classiers, namely RF, SVM-RBF and TFFNet