Posted on 2020-11-06
Machine Learning is a subset of Artificial Intelligence that allows an application to learn from sets of data to function as opposed to the traditional programming wherein an application is programmed to perform functions. In Machine Learning, the algorithm is derived from the output of training data. There are different approaches to train a model in Machine Learning, and it depends on the business problems that it is trying to solve.
There are different categories in machine learning and for this blog, we will just focus on the two:
Supervised Learning starts with sets of data that can be the baseline on how a machine will understand behavior patterns or recognition of objects based on sets of image patterns. An example of the application for this is object recognition, wherein sets of images of an object can already be fed in the training set to recognize what object is being fed in a camera. Supervised Learning, usually requires human intervention for training.
Unsupervised Learning doesn’t start with a set of data and doesn’t require human intervention. This usually applies to processing large set of data. The training comes from iterative collection of data and classify it based on patterns found. Social Media sites such as facebook and Twitter used Unsupervised Learning based on the patterns it found in repetitive collection of data over a period of time to learn the kind of behavior users of the application make.
N-PAX has developed a Proof-of-concept for Object Detection that is intended for Manufacturing industry. Manufacturing companies in the Philippines are mostly operating using conventional means, especially, in the production execution. From inbounding of product parts/components, to sorting, assembling, checking and packing: manual counting and tallying is performed. Personnel scheduling and production particulars are to be scheduled manually. Review of inventory and supplies, production targets, orders, deliveries, among other manufacturing tasks consume time, a number of manpower, energy and money. Thus, an automated monitoring and production execution system may generally decrease the exhaustion of company resources.
For this case study, the application used IP Camera installed in the production site, observing behavior patterns of employees and production output produced in an assembly line. This is an example of supervised learning where the camera was trained to recognize objects through sets of data fed in the application. The data inputted to recognize information are possible objects that may be found in the assembly line. In this case study, a semi-conductor assembly line was used.
The key features of the Object Detection System are as follows:
Motion Detection ER Diagram
Training is the most critical part of any object detection system. However, being the most significant process in the system, it is also the most intricate and laborious part. The more data is fed on the system, the more accurate is its ability to classify and identify the objects being captured by the lenses will become. Thus, the system will become more efficient.
A single object to be accurately detected by the system needs about a minimum of two days training, on-site. Below are some of the training features implemented in this case study.
Raw images from video footages are manually cropped, resized and converted to standard resolution accepted by the program. Each image may be processed at around 2-5 minutes through conventional method. In such case, the minimum training requirement consisting of around 250 images may take a day to complete. And since the process is monotonous and repetitive, it brings boredom and a certain level of physical and mental stress to the person optimizing the images.
After image optimization, this important task will follow. Each image to be used for training must be renamed to avoid overwriting of the images previously stored in the train and test folders. Though this process is done by batch and will take less than a minute to complete, it has to be automated as well so that human interference in the process will be eradicated. Note that this process is within the middle, inattention may halt the succeeding processes.
Soon after renaming image files, the target object within the image must be selected and labeled. This will give each image a class name. For a set of 250 images, manual annotation may take about an hour.
Similar to the preceding process, images are handled to create unique XML file for each. This is done one by one unlike the prior process which can be handled by batch and can be completed in about a minute.
XML files generated from the previous process will then be consolidated and recorded in a CSV file.
Binary data sets are required for training. It is composed of the CSV file generated from the XML files, matched with the original images (with the same file name).
To complete the training, binary datasets will be automatically fed to the system. All the images saved will be the system’s reference in detecting objects that the camera will capture, the bounding box along with the classification label will be displayed when it matches the XML files unique to each image.
The manufacturing industry has many challenges, but with the aid of today’s technological advantages such as machine learning and artificial intelligence, these can be utilized in the production and assembly lines to help reduce labor and production cost and time to market. It is for this reason that N-PAX had embarked on establishing a team focused on Digital Transformation, one area that we can help with is in the manufacturing industry. We can provide consulting services to look into your engineering process, and identify technology touch points where Machine Learning can be implemented.
Email us now (firstname.lastname@example.org) and let’s explore the possibility of Machine Learning.