Computer Vision and the machine learning workflow

The Computer Vision applications with machine learning have a common basic structure. This structure is divided into different steps that are repeated in almost all Computer Vision applications, and some others are omitted. In the following diagram, we show you the different steps involved:

Almost any Computer Vision application starts with a preprocessing stage that is applied to the input image. Preprocessing involveslight removal conditions and noise, thresholding,blur, and so on.

After we apply all the preprocessing steps required to the input image, the second step is segmentation. In thesegmentation step, we need to extract the regions of interest of an image and isolate each one as a unique object of interest. For example,in a face detection system, we need to separate the faces from the rest of the parts in the scene.

After getting the objects inside the image, we continue with the next step. We need toextract all the features of each one detected object; a feature is a vector of characteristics of objects. A characteristic describes our objects and can be thearea of the object, contour, texture pattern, and so on.

Now, we have the descriptor of our object; a descriptor is a feature that describes an object, andwe use these descriptors to train our model or predict one of them. To do this, we need to create a big dataset of features, where hundreds, thousands, and millions of images are preprocessed, and extracted features use all these features in a train model function that we choose:

When we train a dataset, the model learns all the parameters required to predict when a new vector of features with an unknown label is given:

After we get the prediction, sometimes, a post-processing of output data is required;for example, merging multiple classifications todecrease the prediction error or merging multiple labels. A sample case is OCR (Optical Character Recognition),where the classification result is per character, and by combining the results of character recognitions, we construct a word. This means that we can create a postprocessing method to correct errors in detected words.