Though I don't have much idea but you can try with different filter with random 1s and 0s in the matrix. Each filter captures different characteristics of the image which ultimately helps to recognize the words. 

You can start trying to understand what your problem needs.

For example, for a image recognition problem, if you think that a big amount of pixels are necessary for the network recognize the object you will use large filters (as 11x11 or 9x9). If you think what differentiates objects are some small and local features you should use small filters (3x3 or 5x5).

These are some tips, but do not exist any rules, the best way is test some different architectures. If exist papers using CNN for your problem, you can try to use the same architecture and change it. Otherwise, the Alexnet is a good starting point.

Dear Marwa Braiki,

Convolutional Neural Networks are (usually)  supervised methods for image/object recognition. This means that you need to train the CNN using a set of labelled images: this allows to optimize the weights of its convolutional filters, hence learning the filters shape themselsves, to minimize the error.

Once you have decided the size of the filters, as much as the initialization of the filters is important to "guide" the learning, you can indeed initialize them to random values, and let the learning do the work.

Enrico

As Enrico mentioned above, you do not fix the filter coefficients, rather you learn them over several iterations of training. The initialization can be random, or can be based on pre-trained model weights (such as those from the 'modelzoo' in github repos for popular models such as Alexnet, VGG, etc)

Marwa,

I would add to Enrico answer that you can initialize the filters with features learned from unsupervised learning. K-means is a good start and it is a easy algorithm to implement and really fast.

As for size, I believe that the receptive field is what matters most. The receptive field is the result of successive layers of filters stacked on top of each other. Do not forget to account for the pooling layer.

Generally speaking, deeper architecture are better which means that you should favour small filters.

Finally, depending on your machine's memory, sometimes you just need to works with what your machine can handle.

Many thanks for your answers

can you please specify how to exactly loop on training data? I mean, if I have a large training dataset, initialized the filters to random values, ran a forward and a backward propagation .. then how to move to the next epoch ?

use the obtained filters as an initial filters with the next image? as I'm confused, I thought at first there were weights unrelated to the filters(kernels) , then realized that the filters are the weights.

"The initialization ..., or can be based on pre-trained model weights". Does it mean that the kernels of all hidden layers can be initially similar?

Since the hidden layers of a CNN work as trainable feature extractor, for more detailed content based on larger number of pixels shall require bigger filter sizes. But for cases where localized differences are to receive greater attention, smaller filter sizes are required.

ONce you decide the filter size, we randomly initialize the weight of the filter and allow back propagation algorithm to learn weights automatically.

Need more clear explanations.

1. Why so many filters ? The filters used in 1st Conv layer, 2nd Conv layer, and so on ? I am discussing about a problem where a single Remote Sensing image has many objects, and trying to find out a particular object from the image.

2. My 1st question is: How many conv layer is required ? What filters should I use in those conv layers ?

3. What would the architecture of the Fully Connected layers ?

Please help.

Need more clear explanations.

1. Why so many filters ? The filters used in 1st Conv layer, 2nd Conv layer, and so on ? I am discussing about a problem where a single Remote Sensing image has many objects, and trying to find out a particular object from the image.

2. My 1st question is: How many conv layer is required ? What filters should I use in those conv layers ?

3. What would the architecture of the Fully Connected layers ?

Please help.

Vikas Ramachandra , could you please tell me by using which algorithm vgg16 learns filters ?

the optimization is via (batch) gradient descent, for a given loss function (such as binary cross entropy).

Number of filters and filter size are two hyper-parameters, you can tune them during the training. Each dataset requires a different settings of hyper-parameters.

In general, smaller kernel size of odd number are preferred over larger kernel size of even number for better performance of convolutional layer to extract useful features.

I research 3 D volumetric image in CNN. I create pixellabeldatastore for semantic segmentation. But i do not search pixel data for 3 D volumetric.

The choice of a small filter (3X3) corresponds to a good compromise between computational complexity and quality of the resulting image (for example noise reduction while preserving the sharpeness for a median filter, ...)

Dear Marwa Braiki

This article may help you

Article Novel Feature Selection and Voting Classifier Algorithms for...

Marwa Braiki One of the approach is random initialization of filter values which is not necessarily very good. Gaussian distributed random values (so there is some consistent distribution) can be used to initialize filter which will further updated in the training process. Also, there are other method of initializing filter values like He-Norm, Xavier initialization etc look into it.

The question about what should be the size of filter depends upon the problem. Small filters can extract features from smaller neighborhood and large filter for large. Also, have a look at receptive field in Neural Network@.

Bravo!

Can I use only sobel filters(explicitly) in CNN? For detecting edges ?

@ Satish Kumar Satti, you can use filters depending on the task. For example in the space of assisted driving, you can use Sobel filters say to extract the edges of lane boundaries. You may refer to https://www.saama.com/different-kinds-convolutional-filters/ for more details of filters.

Both the size and the number of filters will depend on the complexity of the image and its details. For small and simple images (e.g. Mnist) you would need 3x3 or 5x5 filters and few of them (4, then 8, up to 16) to detect straight lines, curves, obliques, and maybe some color tonality; while for larger images (e.g. faces or landscapes), a bigger receptive field might be needed with 7x7, 9x9 or 11x11 filters, and start with 32 and then 64 and so on to distinguish more shapes and textures. Why use multiples of 2 for the number of filters? Filters are related to feature maps that will be flattened at the end, so it seems to me that it has to do with better RAM organization and batch size work.

After training, are the filters of Conv2D fixed? or do they change?

I'm working on YOLO v4-tiny, and I'm trying to extract the values of the filters used by the Conv2D, but I get different values each time. Can anyone explain that to me?