Application of Convolutional Neural Networks to Achieve Performance Enhancement of Robot Vision Inspection Systems for Industrial Production Lines

Abstract

 In response to the issue of dynamic complexity in robot vision inspection technology on industrial production lines, this paper proposes a method for robot vision inspection that combines convolutional neural networks (CNN) with background difference methods and spatio-temporal context target tracking technology. Using the established image database containing five types of workpieces—bearings, screwdrivers, gears, pliers, wrenches, and other five categories of workpiece images, and utilizing the proposed lightweight residual attention and decoupled attention mechanisms to design the FCN8s model. Experimental validation revealed that the model achieved an average pixel accuracy, average accuracy, and average IoU accuracy of 98.30%, 82.46%, and 73.62%, respectively, on the test set, meeting the real-time requirements of industrial visual inspection. In terms of accuracy, the recognition rate increased by approximately 15.97% compared to traditional methods, and the speed increased by approximately 83.26% to 84.65%. This paper adopts a multi-technology fusion solution to provide a more reliable and practical means for target detection and recognition in complex industrial environments, thereby providing a robot vision system for intelligent manufacturing.