Deep learning: How OpenCV’s blobFromImage works

Today’s blog post is inspired by a number of PyImageSearch readers who have commented on previous deep learning tutorials wanting to understand what exactly OpenCV’s blobFromImage  function is doing under the hood.

You see, to obtain (correct) predictions from deep neural networks you first need to preprocess your data.

In the context of deep learning and image classification, these preprocessing tasks normally involve:

  1. Mean subtraction
  2. Scaling by some factor

OpenCV’s new deep neural network ( dnn ) module contains two functions that can be used for preprocessing images and preparing them for classification via pre-trained deep learning models.

In today’s blog post we are going to take apart OpenCV’s cv2.dnn.blobFromImage  and cv2.dnn.blobFromImages  preprocessing functions and understand how they work.

To learn more about image preprocessing for deep learning via OpenCV,  just keep reading.

Looking for the source code to this post?
Jump right to the downloads section.

Deep learning: How OpenCV’s blobFromImage works

OpenCV provides two functions to facilitate image preprocessing for deep learning classification:

  • cv2.dnn.blobFromImage
  • cv2.dnn.blobFromImages

These two functions perform

  1. Mean subtraction
  2. Scaling
  3. And optionally channel swapping

In the remainder of this tutorial we’ll:

  1. Explore mean subtraction and scaling
  2. Examine the function signature of each deep learning preprocessing function
  3. Study these methods in detail
  4. And finally, apply OpenCV’s deep learning functions to a set of input images

Let’s go ahead and get started.

Deep learning and mean subtraction

Figure 1: A visual representation of mean subtraction where the RGB mean (center) has been calculated from a dataset of images and subtracted from the original image (left) resulting in the output image (right).

Before we dive into an explanation of OpenCV’s deep learning preprocessing functions, we first need to understand mean subtraction. Mean subtraction is used to help combat illumination changes in the input images in our dataset. We can therefore view mean subtraction as a technique used to aid our Convolutional Neural Networks.

Before we even begin training our deep neural network, we first compute the average pixel intensity across all images in the training set for each of the Red, Green, and Blue channels.

This implies that we end up with three variables:

\mu_R\mu_G, and \mu_B

Typically the resulting values are a 3-tuple consisting of the mean of the Red, Green, and Blue channels, respectively.

For example, the mean values for the ImageNet training set are R=103.93, G=116.77, and B=123.68 (you may have already encountered these values before if you have used a network that was pre-trained on  ImageNet).

However, in some cases the mean Red, Green, and Blue values may be computed channel-wise rather than pixel-wise, resulting in an MxN matrix. In this case the MxN matrix for each channel is then subtracted from the input image during training/testing.

Both methods are perfectly valid forms of mean subtraction; however, we tend to see the pixel-wise version used more often, especially for larger datasets.

When we are ready to pass an image through our network (whether for training or testing), we subtract the mean, \mu, from each input channel of the input image:

R = R - \mu_R

G = G - \mu_G

B = B - \mu_B

We may also have a scaling factor, \sigma, which adds in a normalization:

R = (R - \mu_R) / \sigma

G = (G - \mu_G) / \sigma

B = (B - \mu_B) / \sigma

The value of \sigma may be the standard deviation across the training set (thereby turning the preprocessing step into a standard score/z-score). However,\sigma may also be manually set (versus calculated) to scale the input image space into a particular range — it really depends on the architecture, how the network was trained, and the techniques the implementing author is familiar with.

It’s important to note that not all deep learning architectures perform mean subtraction and scaling! Before you preprocess your images, be sure to read the relevant publication/documentation for the deep neural network you are using.

As you’ll find on your deep learning journey, some architectures perform mean subtraction only (thereby setting \sigma=1). Other architectures perform both mean subtraction and scaling. Even other architectures choose to perform no mean subtraction or scaling. Always check the relevant publication you are implementing/using to verify the techniques the author is using.

Mean subtraction, scaling, and normalization are covered in more detail inside Deep Learning for Computer Vision with Python.

OpenCV’s blobFromImage and blobFromImages function

Let’s start off by referring to the official OpenCV documentation for cv2.dnn.blobFromImage :

[blobFromImage] creates 4-dimensional blob from image. Optionally resizes and crops image from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels.

Informally, a blob is just a (potentially collection) of image(s) with the same spatial dimensions (i.e., width and height), same depth (number of channels), that have all be preprocessed in the same manner.

The  cv2.dnn.blobFromImage  and  cv2.dnn.blobFromImages  functions are near identical.

Let’s start with examining the  cv2.dnn.blobFromImage  function signature below:

blob = cv2.dnn.blobFromImage(image, scalefactor=1.0, size, mean, swapRB=True)

I’ve provided a discussion of each parameter below:

  1. image : This is the input image we want to preprocess before passing it through our deep neural network for classification.
  2. scalefactor : After we perform mean subtraction we can optionally scale our images by some factor. This value defaults to 1.0 (i.e., no scaling) but we can supply another value as well. It’s also important to note that scalefactor  should be 1 / \sigma as we’re actually multiplying the input channels (after mean subtraction) by scalefactor .
  3. size : Here we supply the spatial size that the Convolutional Neural Network expects. For most current state-of-the-art neural networks this is either 224×224, 227×227, or 299×299.
  4. mean : These are our mean subtraction values. They can be a 3-tuple of the RGB means or they can be a single value in which case the supplied value is subtracted from every channel of the image. If you’re performing mean subtraction, ensure you supply the 3-tuple in (R, G, B) order, especially when utilizing the default behavior of swapRB=True .
  5. swapRB : OpenCV assumes images are in BGR channel order; however, the mean value assumes we are using RGB order. To resolve this discrepancy we can swap the R and B channels in image  by setting this value to True. By default OpenCV performs this channel swapping for us.

The  cv2.dnn.blobFromImage  function returns a  blob  which is our input image after mean subtraction, normalizing, and channel swapping.

The  cv2.dnn.blobFromImages  function is exactly the same:

blob = cv2.dnn.blobFromImages(images, scalefactor=1.0, size, mean, swapRB=True)

The only exception is that we can pass in multiple images, enabling us to batch process a set of  images .

If you’re processing multiple images/frames, be sure to use the  cv2.dnn.blobFromImages  function as there is less function call overhead and you’ll be able to batch process the images/frames faster.

Deep learning with OpenCV’s blobFromImage function

Now that we’ve studied both the blobFromImage  and blobFromImages  functions, let’s apply them to a few example images and then pass them through a Convolutional Neural Network for classification.

As a prerequisite, you need OpenCV version 3.3.0 at a minimum. NumPy is a dependency of OpenCV’s Python bindings and imutils is my package of convenience functions available on GitHub and in the Python Package Index.

If you haven’t installed OpenCV, you’ll want to follow the latest tutorials available here, and be sure to specify OpenCV 3.3.0 or higher when you clone/download opencv  and opencv_contrib .

The imutils package can be installed via pip :

Assuming your image processing environment is ready to go, let’s open up a new file, name it blob_from_images.py , and insert the following code:

First we import imutils , numpy , and cv2  (Lines 2-4).

Then we read synset_words.txt  (the ImageNet Class labels) and extract classes , our class labels, on Lines 7 and 8.

To load our model model from disk we use the DNN function,  cv2.dnn.readNetFromCaffe , and specify bvlc_googlenet.prototxt  as the filename parameter and bvlc_googlenet.caffemodel  as the actual model file (Lines 11 and 12).

Note: You can grab the pre-trained Convolutional Neural Network, class labels text file, source code, and example images to this post using the “Downloads” section at the bottom of this tutorial.

Finally, we grab the paths to the input images on Line 15. If you’re using Windows you should change the path separator here to ensure you can correctly load the image paths.

Next, we’ll load images from disk and pre-process them using blobFromImage :

In this block, we first load the image  (Line 20) and then resize it to 224×224 (Line 21), the required input image dimensions for GoogLeNet.

Now we’re to the crux of this post.

On Line 22, we call cv2.dnn.blobFromImage  which, as stated in the previous section, will create a 4-dimensional blob  for use in our neural net.

Let’s print the shape of our blob so we can analyze it in the terminal later (Line 23).

Next, we’ll feed blob  through GoogLeNet:

If you’re familiar with recent deep learning posts on this blog, the above lines should look familiar.

We feed the blob  through the network (Lines 28 and 29) and grab the predictions, preds .

Then we sort preds  (Line 33) with the most confident predictions at the front of the list, and generate a label text to display on the image. The label text consists of the class label and the prediction percentage value for the top prediction (Lines 34 and 35).

From there,  we write the label text  at the top of the image  (Lines 36 and 37) followed by displaying the image  on the screen and waiting for a keypress before moving on (Lines 40 and 41).

Now it’s time to use the plural form of the blobFromImage  function.

Here we’ll do (nearly) the same thing, except we’ll instead create and populate a list of images  followed by passing the list as a parameter to blobFromImages :

First we initialize our images  list (Line 44), and then, using the imagePaths , we read, resize, and append the image  to the list (Lines 49-52).

Using list slicing, we’ve omitted the first image from imagePaths  on Line 49.

From there, we pass the images  into cv2.dnn.blobFromImages  as the first parameter on Line 55. All other parameters to cv2.dnn.blobFromImages  are identical to cv2.dnn.blobFromImage  above.

For analysis later we print blob.shape  on Line 56.

We’ll next pass the blob  through GoogLeNet and write the class label and prediction at the top of each image:

The remaining code is essentially the same as above, only our for  loop now handles looping through each of the imagePaths  (again, omitting the first one as we have already classified it).

And that’s it! Let’s see the script in action in the next section.

OpenCV blobfromImage and blobFromImages results

Now we’ve reached the fun part.

Go ahead and use the “Downloads” section of this blog post to download the source code, example images, and pre-trained neural network. You will need the additional files in order to execute the code.

From there, fire up a terminal and run the following command:

The first terminal output is with respect to the first image found in the images  folder where we apply the cv2.dnn.blobFromImage  function:

The resulting beer glass image is displayed on the screen:

Figure 2: An enticing beer has been labeled and recognized with high confidence by GoogLeNet. The blob dimensions resulting from blobFromImage are displayed in the terminal.

That full beer glass makes me thirsty. But before I enjoy a beer myself, I’ll explain why the shape of the blob is (1, 3, 224, 224) .

The resulting tuple has the following format:

(num_images=1, num_channels=3, width=224, height=224)

Since we’ve only processed one image, we only have one entry in our blob . The channel count is three for BGR channels. And finally 224×224 is the spatial width and height for our input image.

Next, let’s build a blob  from the remaining four input images.

The second blob’s shape is:

Since this blob contains 4 images, the num_images=4  . The remaining dimensions are the same as the first, single image, blob.

I’ve included a sample of correctly classified images below:

Figure 3: My keyboard has been correctly identified by GoogLeNet with a prediction confidence of 81%.

Figure 4: I tested the pre-trained network on my computer monitor as well. Here we can see the input image is correctly classified using our Convolutional Neural Network.

Figure 5: A NASA space shuttle is recognized with a prediction value of over 99% by our deep neural network.

Summary

In today’s tutorial we examined OpenCV’s blobFromImage  and blobFromImages  deep learning functions.

These methods are used to prepare input images for classification via pre-trained deep learning models.

Both blobFromImage  and blobFromImages  perform mean subtraction and scaling. We can also swap the Red and Blue channels of the image depending on channel ordering. Nearly all state-of-the-art deep learning models perform mean subtraction and scaling — the benefit here is that OpenCV makes these preprocessing tasks dead simple.

If you’re interested in studying deep learning in more detail, be sure to take a look at my brand new book, Deep Learning for Computer Vision with Python.

Inside the book you’ll discover:

  • Super practical walkthroughs that present solutions to actual, real-world image classification problems, challenges, and competitions.
  • Detailed, thorough experiments (with highly documented code) enabling you to reproduce state-of-the-art results.
  • My favorite “best practices” to improve network accuracy. These techniques alone will save you enough time to pay for the book multiple times over.
  • ..and much more!

Sound good?

Click here to start your journey to deep learning mastery.

Otherwise, be sure to enter your email address in the form below to be notified when future deep learning tutorials are published here on the PyImageSearch blog.

See you next week!

Downloads:

If you would like to download the code and images used in this post, please enter your email address in the form below. Not only will you get a .zip of the code, I’ll also send you a FREE 11-page Resource Guide on Computer Vision and Image Search Engines, including exclusive techniques that I don’t post on this blog! Sound good? If so, enter your email address and I’ll send you the code immediately!

, , , ,

24 Responses to Deep learning: How OpenCV’s blobFromImage works

  1. Pure November 6, 2017 at 11:32 am #

    Email notification squad, where you at!

    Hey Adrian, I’m really enjoying your series of articles on deep learning. I love the powerful technology we can create by tying in computer vision and neural networks… It really is the combination that allows us to make *magic* software that makes people say “Wow!”

    Anyways, this is a bit off topic, but I was wondering if you would be so kind as to write an article on making a “people counter” with OpenCV — that is, a program that counts people going in and out of a building via a live webcam feed. There are no great resources available online for this, so if you would write one I’m sure it would drive plenty of traffic to your site. It’s a win win for both of us!

    That being said, I love all the content you’re putting out now. Keep doing your thing 🙂

    Take care!

    • Adrian Rosebrock November 6, 2017 at 4:22 pm #

      Sure, I can absolutely do a blog post on that. I’m right in the middle of a deep learning series, but I’ll add it to the idea queue. Thanks for the suggestion.

  2. Maxim November 6, 2017 at 12:12 pm #

    Hi

    Thank for a new post. I tried to experiment with dnn modul of opencv for semantic segmentation tasks but I had to refuse from it. Maybe I just didnt find a suitable network. although, later on I didnt find a suitable solution and with using original frameworks. (or quality is poor or size of pretrained net is too huge). Maybe you, Adrian, have some suggestion what network can be used in mobile devices for semantic segmentation tasks?

    • Adrian Rosebrock November 6, 2017 at 4:22 pm #

      I would suggest doing some research on MobileNet and then trying out MobileNet + semantic segmentation.

      • Maxim November 7, 2017 at 6:08 am #

        MobileNet is attended for classifications. Models which made on base of MobileNet also is not perfect unfortunately. UNet is too big. Pretrained caffe model what I found is 124Mb and it is not suitable for mobile devices. PSPNet is about 30Mb, it is better, but quality is poor. Even their web demo on site is not work well on arbitrary images. It is a common problem of all segmentation networks. They work fine on images on which these were trained (i.e. from cityscapes set) but on any random photo results are too far from desirable. I found only one network what works moreless fine on random images – it is Sharpnet by facebook. But their size is more 500Mb 🙁
        Ideally, it was needed to combine i.e. ENet + SharpNet and make a compact net, but I didnt find this already done. To make it self – this task will require to drop all other tasks and devote all time to studying NN more deeply. And the most defensively is what so kind of already done mobile network will be available 1-2 year later, but not now yet.

        • Adrian Rosebrock November 9, 2017 at 7:00 am #

          MobileNet by itself is used for image classification. MobileNet + SSD can become an object detector. MobileNet can also be combined with a segmentation framework as well.

          You also mentioned “they work fine on images they were trained on” which is actually what all machine learning algorithms do. There is no such thing as a “perfect generalizability” to images a network were not trained on, especially if your input images dramatically vary from the training set.

          Deep learning-based segmentation will get better for sure, but that also implies that our datasets need to become more robust as well.

          • Maxim November 9, 2017 at 10:07 am #

            Good advise. I’ll try to find some project with pretrained model on base of MobileNet. (havent yet found)
            Regarding “perfect generalizability” – when I’ve becomes acquainted with MachineLearning then for task of classification and object detection it looks like a miracle, and usually it works fine on random pictures. And only with a segmentation task I’ve met a disappointment. Of course, it is more complicated task than previous two but a faith in a miracle dead last. Even monstrous networks like SharpMask doesnt see stable results. They are interested for experiments but not yet applicable for real tasks. During google searching I’ve met several commercial projects what promise a good result and appropriate sizes for mobile solution. But they havent demo versions and I cannot estimate do they say true or it is just advertising promises. But they motivate to continue to experiment with existed open source projects

  3. Tuan November 7, 2017 at 1:08 am #

    Can you write a post that introduce deep learning feature on OpenCV 3.3.0? Thank article!

    • Adrian Rosebrock November 9, 2017 at 7:05 am #

      Hi Tuan — I have already done this. Please see this post.

  4. chintan zaveri November 7, 2017 at 2:14 am #

    Hey,
    Thanks for amazing tutorial. Is there any help available for image segmentation using dnn module of openCV?

  5. Mustafa Qamar-ud-Din November 7, 2017 at 6:18 am #

    Thank you for the very informative blog & newsletter. I was wondering whether you can advise necessary tools for applying these techniques in production environment around RESTful APIs 😉

  6. loukas November 7, 2017 at 7:48 am #

    very good job!!! i want to ask if that can work on real time object detection??

    • Adrian Rosebrock November 9, 2017 at 6:54 am #

      Absolutely. Please see this blog post.

      • loukas November 14, 2017 at 3:05 pm #

        thnx for your answer but on this link you have only 20 model. how i can import this model!!! thnx again for your time

        • Adrian Rosebrock November 15, 2017 at 1:01 pm #

          The object detection model supports the 20 COCO classes. The image classification model was trained on the 1,000 ImageNet labels. You cannot take a model trained for image classification and use it for object detection. You would instead need to train your object detection model from scratch OR apply transfer learning via fine-tuning. For what it’s worth, I’m covering object detection in detail inside Deep Learning for Computer Vision with Python.

  7. PBS November 7, 2017 at 11:35 am #

    Hey Adrian

    For me, a blob is a set of connected pixels, or a connected component, usually found in binary images. Why did they choose “blob” for this operation, which seems like has nothing to do with traditional blob?

    • Adrian Rosebrock November 9, 2017 at 6:53 am #

      Calling the output of these functions a “blob” is a decision by the OpenCV developers — I had nothing to do with this choice. I’m not sure why they choose the name blob, I suppose you would need to ask them.

  8. Tuan November 8, 2017 at 5:00 am #

    Hi Adrian,
    How to “rectangle” for detected object? Based on your code, I could not do that! (If I use MobileNet module, it can help me but less accurate)

    • Adrian Rosebrock November 9, 2017 at 6:37 am #

      I cover how to compute the bounding box rectangle for a given object in this blog post. You cannot take a model that was trained for image classification and use it for object detection. Object detection deep learning models follow a specific framework and need to be trained in a very specific way. I’m covering object detection deep learning models inside my book, Deep Learning for Computer Vision with Python.

      • Tuan November 16, 2017 at 11:46 pm #

        I am interested in your book and your website. But i have no money to buy it. Hope you successfully!

  9. zz November 9, 2017 at 5:39 am #

    good

  10. Fábio Uechi November 15, 2017 at 10:07 pm #

    Hi Adrian,

    Thanks for the very informative post.
    I have a question regarding mean subtraction.
    I’m trying to load another model published in the caffe model zoo (https://github.com/BVLC/caffe/wiki/Model-Zoo#models-for-age-and-gender-classification).
    Apart from the .prototxt and .caffemodel files it also provides a mean.binaryproto file. Do you happen to know how can we figure out the mean and scalefactor parameters values from this file ?

    • Adrian Rosebrock November 18, 2017 at 8:25 am #

      Hi Fábio — you’ll need to use the “caffe” Python bindings (which will require you to install and compile Caffe). Something like this would be a step in the right direction.

Leave a Reply