How to Implement the Data Science Lifecycle: Computer Vision

Data science and the data science lifecycle is a methodology that provides a structured framework for projects dealing with very large amounts of information. With the increasing popularity of big data, many industries are experiencing rapid growth rates in the amount of data they generate.

Much of this information can be useful for identifying repetitive failures in a process and opportunities for improvement. A company facing one of these scenarios can find it justifiable to launch a big data project. Working with big data in an organized and efficient manner requires adherence to the data science lifecycle methodology.

 

Figure 1. The iterative six phases of the CRISP-DM model are shown. Image used courtesy of Kenneth Jensen

 

There are numerous instances and variations of the data science lifecycle model that have been published over several years. Out of all of them, the Cross-Industry Standard Process for Data Mining (CRISP-DM) model stands out as one of the first ones created, as well as one of the most popular. The applications reviewed in this article are based on CRISP-DM. 

The CRISP-DM model consists of six phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. This sequence helps to guide a team through the successful execution of a big data project that seeks to solve a problem within a process.

 

What is Computer Vision?

Computer vision is among one of the fastest-growing engineering fields. It is concerned with how computers can understand their surroundings by analyzing digital images and videos. It has become well-known in recent years thanks to applications like facial recognition and autonomous vehicles.

 

Figure 2. Computer vision is essential for self-learning machines and big data. Image used courtesy of Cognex

 

Manufacturing processes and, in general, industrial settings, have also seen increased penetration of computer vision systems. Some specific computer vision applications in the industry are object detection, classification, tagging, semantic segmentation, and instance segmentation.

 

Using CRISP-DM in Computer Vision

Modern computer vision systems are perfect cases for applying the CRISP-DM model. Analyzing large volumes of high-quality images and videos requires vast computer processing power. Therefore, it is important to organize and clean the data pool as much as possible to optimize computer resources. The next paragraphs review the six phases of the CRISP-DM model applied to computer vision projects.

 

Understanding and Preparing the Data

In the business understanding phase, much like any other application, the problem is defined, and the goals and metrics for success are established. The goals of a computer vision project are clear, concise, and in line with the capabilities of the vision system available. Some examples of goals include classification, tagging, segmentation, action recognition, and video tracking.

The next phase, data understanding, is crucial. This is when parameters for the image or video set’s quantity, diversity, accuracy, and quality are defined. It is important to use as much visual data as possible for a meaningful result, as long as the computer performance is not significantly compromised.

Diversity of visual data helps the vision equipment understand as many different external conditions as possible. Examples of these external conditions include lighting, orientations, and colors. Diversity also refers to the quality of the visual data because it helps to have low- and high-resolution imagery for comparison between them.

Accuracy is perhaps the most critical criterion in this case. The vision system must be taught in advance how to interpret visual data and, therefore, the training samples used must be accurate. Quality deals with the resolution of the images, which nowadays, nearly all new vision systems offer high-resolution.

Data samples are reviewed for completeness and meaningfulness during the data preparation phase. Also, the data are labeled and classified for correct decision-making in the last phase.

Data labeling is a very important task in computer vision systems. This refers to what the computer interprets the content of a single frame in an image to be, thus what kind of entities are identified. In this CRISP-DM model phase, it is crucial to verify that the visual data analyzed in the previous phase has been correctly identified and labeled; otherwise, it can lead to erroneous results.

 

Modeling and Evaluating the Data

Modeling is the phase when software tools and algorithms are selected. Python is among the preferred programming languages for coding visual data models. The neural network algorithm is among the most popular for deep learning in vision systems. The neural network is a machine learning methodology that attempts to emulate how the human brain functions and apply that to the software tool.

The basic process in neural networks consists of receiving data, looking in the neural network for patterns, assimilating the new data into the network, and outputting a result of the evaluation.

 

Figure 3. Neural networks are an important type of algorithm used in computer vision. Image used courtesy of Matt Britt

 

In the evaluation phase, the outcomes of running the visual data through the model are measured and assessed against a previously defined metric or key performance indicator (KPI). Accuracy and precision are two important metrics defined by a ratio of correct vision system interpretations versus the total number of tests.

If the evaluation produces too many inaccuracies, the iterative model suggests returning to the business understanding phase to rethink the problem and the project objectives.

The last phase is deployment—when the solution is migrated to a live production system. This is the right time to make considerations of cost and scalability. Security also becomes important because now the visual data will be moved to an environment that likely has more users and possible points of breaching. The deployment phase is iterative because once the go-live is completed, there is a period of permanent monitoring and solution upgrades.

Computer vision, continuing to rise in utilization, is highly complex and needs large amounts of data to properly train the chosen AI model. Implementing the CRISP-DM data science lifecycle model can help keep plans on track.