List of Operators ↓
This chapter explains how to use classification based on deep learning, both for the training and inference phases.
Classification based on deep learning is a method, in which an image gets a set of confidence values assigned. These confidence values indicate how likely the image belongs to each of the distinguished classes. Thus, if we regard only the top prediction, classification means to assign a specific class out of a given set of classes to an image. This is illustrated with the following schema.
In order to do your specific task, thus to classify your data into the classes you want to have distinguished, the classifier has to be trained accordingly. In HALCON, we use a technique called transfer learning (see also the chapter Deep Learning). Hence, we provide pretrained networks, representing classifiers which have been trained on huge amounts of labeled image data. These classifiers have been trained and tested to perform well on industrial image classification tasks. One of these classifiers, already trained for general classifications, is now retrained for your specific task. For this, the classifier needs to know, which classes are to be distinguished and how such examples look like. This is represented by your dataset, i.e., your images with the corresponding ground truth labels. More information on the data requirements can be found in the section “Data for classification”.
For the specific system requirements in order to apply deep learning
classification, please refer to the HALCON
Have a look at the HDevelop example
for a short-and-simple overview and
classify_pill_defects_deep_learning.hdev for a more sophisticated
workflow (both can be found under
They provide great guidance on how the different parts can be used
First, a pretrained network has to be read using the operator
This operator is used as well when you want to read your own trained
networks, after you saved them with
To read the data for your deep learning classification training the procedure
is available. Using this procedure you can get a list of image file paths and their respective labels (the ground truth labels) as well as a list of the unique classes, to which at least one of the listed images belongs.
The network will impose several requirements on the images,
as the image dimensions and the gray value range.
The default values are listed in
. These are
the values with which the networks have been pretrained.
The network architectures allow different image dimensions, which can
be set with
, but depending on the
network a change may make a retraining necessary.
The actually set values can be retrieved with
preprocess_dl_classifier_images provides great
guidance on how to implement such a preprocessing stage.
We recommend to preprocess and store all images used for the training
before starting the classifier training, since this speeds up the
Next, we recommend to split the dataset into three distinct datasets which are used for training, validation, and testing, see the section “Data” in the chapter Deep Learning. This can be achieved using the procedure
You need to specify the
'classes' (determined before by
read_dl_classifier_data_set) you want to differentiate
with your classifier. For this, the operator
This operator can also be used to set hyperparameters, which are
important for training, e.g.
For a more detailed explanation, see the chapter
Deep Learning and the documentation of
Once your network is set up and your data prepared it is time to train the classifier for your specific task.
Set the hyperparameters used for training with the operator
For an overview of possible hyperparameters, see the documentation of
. Additional explanations can be found
in the chapter Deep Learning.
To train the classifier the operator
is available. The intermediate training results are stored in the output handle.
As the name of
operator processes a batch of data (images and ground truth labels) at
We iterate through our training data in order to train
the classifier successively with
You can repeat this process multiple times and iterate over so many
training epochs until you are satisfied with the training result.
To know how well the classifier learns the new task, the procedure
is provided. With it you can plot the classification errors during training. To compute the input necessary for the visualization, the procedures
are available. With them you can reduce the number of images used
for this classification validation, apply the classifier on the
selected data, and compute, for example, the top-1 error.
Have a look at the HDevelop example
classify_pill_defects_deep_learning.hdev to see how these
procedures can be used together.
Your classifier is trained for your task and ready to be applied. But before deploying in the real world you should evaluate how well the classifier performs on basis of your test data.
To apply the classifier on a set containing an arbitrary number of images, use the operator
The runtime of this operator depends on the number of batches needed for the given image set.
The results are returned in a handle.
To retrieve the predicted classes and confidences, use the operator
Now it is time to evaluate these results. The performance of the
classifier can be evaluated as during the training with
To visualize and analyze the classifier quality, the confusion matrix is a useful tool (see below for an explanation). For this, you can use the procedures
The interactive procedure gives you the possibility to select examples of a specific category, but it does not work with exported code.
Additionally, after applying the classifier on a set of data, you can use the procedure
to display and return images according to certain criteria, e.g., wrongly classified ones. Then, you might want to use this input for the procedure
to display a heatmap of the input image, with which you can analyze which regions of the image are relevant for the classification result.
When your classifier is trained and you are satisfied with its performance, you can use it to classify new images. For this, you simply preprocess your images according to the network requirements (i.e., the same way as you did for your dataset used for training the classifier) and apply the classifier using
We distinguish between data used for training and data for inference. Latter one consists of bare images. But for the former one you already know to which class the images belong and provide this information over the corresponding labels.
The training data is used to train a classifier for your specific task.
With the aid of this data the classifier can learn which classes are to be
distinguished and how their representatives look like.
In classification, the image is classified as a whole.
Therefore, the training data consists of images and their ground truth
labels, thus the class you say this image belongs to.
Note that the images should be as representative as possible for your
There are different possible ways, how to store and retrieve the ground
read_dl_classifier_data_set supports the following
sources of the ground truth label for an image:
The last folder name containing the image
The file name.
For training a classifier, we use a technique called transfer learning (see the chapter Deep Learning). For this, you need less resources, but still a suitable set of data which is generally in the order of hundreds to thousands per class. While in general the network should be more reliable when trained on a larger dataset, the amount of data needed for training also depends on the complexity of the task. You also want enough training data to split it into three subsets, which are preferably independent and identically distributed, see the section “Data” in the chapter Deep Learning.
Regardless of the application,
the network poses requirements on the images regarding the image
dimensions, the gray value range, and the type.
The specific values depend on the network itself and can be queried with
You can find guidance on how to implement such a preprocessing stage
by the procedure
When we classify an image, we obtain a set of confidence values, telling us the affinity of the image to every class. It is also possible to compute the following values.
In classification whole images are classified. As a consequence, the instances of a confusion matrix are images. See the chapter Deep Learning for explanations on confusion matrices.
You can generate a confusion matrix with the aid of the procedures
Thereby, the interactive procedure gives you the possibility to select
examples of a specific category, but it does not work with exported code.
From such a confusion matrix you can derive various values. The precision is the proportion of all correct predicted positives to all predicted positives (true and false ones). Thus, it is a measure of how many positive predictions really belong to the selected class.
The recall, also called the "true positive rate", is the proportion of all correct predicted positives to all real positives. Thus, it is a measure of how many samples belonging to the selected class were predicted correctly as positives.
A classifier with high recall but low precision finds most members of positives (thus members of the class), but at the cost of also classifying many negatives as member of the class. A classifier with high precision but low recall is just the opposite, classifying only few samples as positives, but most of these predictions are correct. An ideal classifier with high precision and high recall will classify many samples as positive with a high accuracy.
To represent this with a single number, we compute the F1-score, the harmonic mean of precision and recall. Thus, it is a measure of the classifier's accuracy.
For the example from the confusion matrix shown in Deep Learning we get for the class 'apple' the values precision: 1.00 (= 68/(68+0+0)), recall: 0.74 (= 68/(68+21+3)), and F1-score: 0.85 (=2*(1.00*0.74)/(1.00+0.74)).