Research On Objective Visual Detection Method In Complex Industrial Scenes

Visual inspection of mechanical components in industrial scenes is one of the important technological cores in the field of smart manufacturing.With the advent of a new era of industrial intelligent manufacturing,the performance of visual inspection of component parts in industrial scenes is becoming increasingly important for improving the overall level of smart manufacturing.Compared to the common visual detection of fixed light sources and intangible variables in the industrial scene,there is an urgent need in the manufacturing field for visual inspection technology to be applied to more complex industrial scenes.However,the inherent change of the light source in the complex industrial scene and the deformation of the detected target cause the visual inspection task of the mechanical component target in the scene to be extremely difficult.In this paper,the common mechanical parts targets in complex industrial scenes are visual inspection objects,and the visual detection methods based on statistical machine learning and deep learning algorithms are the research contents.The research aims are to achieve the goals of single or multiple types of parts and components in complex industrial scenarios.Visual identification and location detection tasks.The main research work of this paper is as follows:1.Established three types of image data sets suitable for visual inspection tasks in industrial scenes.(1)1008 frame images of single target gear detection data set,including positive sample image and negative sample image;(2)Image level identification classification experiment data set A total of 750 frame images containing bearings,screwdrivers,gears,pliers,and spanners.A total of 2248 frames of mechanical parts and tools are used.Each frame of image contains only the rectangular area where the target is located.(3)A multi-component target detection experiment data set at the regional level.The same five targets as the second class,but each frame contains the target and background.2.A single target gear vision detection method based on joint model matching is proposed.For specific applications in industrial scenes,the image enhancement technology is used to enhance the gradient information of the image,so that the target features have better separability;and then the joint vision detection is performed by using the whole support vector machine model and the local combined support vector regression model.The goal of visual detection of gear targets in industrial scenes has been achieved and detection accuracy has been improved.3.Based on the matconvnet deep learning framework,a convolutional neural network was designed and built for the visual classification task of the components contained in the entire image.The goal classification experiments of five mechanical parts of the image level were achieved,and 250 frames of test images were also performed.The experimental results were quantitatively evaluated and analyzed.4.Fast R-CNN-based regional convolutional neural network multi-component target classification,identification and location detection.The experimental results of the target recognition and positioning of 1124 frame test images for five types of mechanical parts in an unstructured industrial scene show that: In this paper,the target detection methods for many types of parts and components in industrial scenes have achieved good detection results.The average classification accuracy and recall rate have reached 99.4% and 99.5%,respectively,and the average positioning accuracy has reached 89.5%.This paper establishes visual inspection datasets in three kinds of industrial scenes,and successfully implements visual inspection of industrial scenes based on shallow learning and deep learning.The target detection of mechanical components in complex industrial scenes has broad application prospects in the field of smart manufacturing.This work has conducted in-depth research on existing detection algorithms,and provides reference for the application of visual detection algorithms to the field of smart manufacturing.