A feasibility study on the use of agent-based image recognition on a desktop computer for the purpose of quality control in a production environment

Abstract

A multi-threaded, multi-agent image recognition software application called RecMaster has been developed specifically for the purpose of quality control in a production environment. This entails using the system as a monitor to identify invalid objects moving on a conveyor belt and to pass on the relevant information to an attached device, such as a robotic arm, which will remove the invalid object.

The main purpose of developing this system was to prove that a desktop computer could run an image recognition system efficiently, without the need for high-end, high-cost, specialised computer hardware. The programme operates by assigning each agent a task in the recognition process and then waiting for resources to become available. Tasks related to edge detection, colour inversion, image binarisation and perimeter determination were assigned to individual agents.

Each agent is loaded onto its own processing thread, with some of the agents delegating their subtasks to other processing threads. This enables the application to utilise the available system resources more efficiently.

The application is very limited in its scope, as it requires a uniform image background as well as little to no variance in camera zoom levels and object to lens distance. This study focused solely on the development of the application software, and not on the setting up of the actual imaging hardware. The imaging device, on which the system was tested, was a web cam capable of a 640 x 480 resolution. As such, all image capture and processing was done on images with a horizontal resolution of 640 pixels and a vertical resolution of 480 pixels, so as not to distort image quality. The application locates objects on an image feed - which can be in the format of a still image, a video file or a camera feed - and compares these objects to a model of the object that was created previously. The coordinates of the object are calculated and translated into coordinates on the conveyor system. These coordinates are then passed on to an external recipient, such as a robotic arm, via a serial link.

The system has been applied to the model of a DVD, and tested against a variety of similar and dissimilar objects to determine its accuracy. The tests were run on both an AMD- and Intel-based desktop computer system, with the results indicating that both systems are capable of efficiently running the application. On average, the AMD-based system tended to be 81% faster at matching objects in still images, and 100% faster at matching objects in moving images. The system made matches within an average time frame of 250 ms, making the process fast enough to be used on an actual conveyor system. On still images, the results showed an 87% success rate for the AMD-based system, and 73% for Intel. For moving images, however, both systems showed a 100% success rate.