Research On Screw Comprehensive Detection Method Based On Machine Vision And Deep Learning

At present,China's large manufacturing companies have started the wave of industrial transformation,and the demand for high-quality industrial parts in various industries is increasing day by day.As an indispensable part of industrial equipment,the quality of screws has a great influence on the performance of industrial equipment.With the development of computer and artificial intelligence technology,machine vision and deep learning technology have been widely used in industrial production.At the same time,the method of large-scale automated production of equipment under industrialization is becoming more and more common.The comprehensive detection method of screw size and defect characteristics based on non-contact mode has become a hot spot in industry research.Since defect features are difficult to be quantified and extracted,it is more difficult to use machine vision technology to detect defects on the industrial assembly line,with a higher rate of false positives and more complicated processing.In recent years,network models based on deep learning have achieved great success on all major datasets,proving that technology can achieve good results in the field of image processing.This also provides us with a faster and more accurate method for industrial intelligent detection.Based on the existing theory,this paper analyzes and studies the important characteristics of screws used in industry,and proposes a comprehensive inspection method for screws based on machine vision and deep learning.This method focuses on the dimensional features and defect features of screws.The research content is mainly divided into the following four aspects:1)Construction of visual inspection system: First,based on the principle of projection method,build the hardware platform of the optical system,choose CCD industrial camera,and use the same level of the optical lens as industrial camera pixels.Then adjust the intensity of the LED parallel light source through the light source controller,select the appropriate lighting method to collect the screw image.Finally,according to the correspondence between different coordinate systems,the principle of transformation is analyzed to eliminate the influence of distortion on the image during image acquisition,and the entire visual system is calibrated.2)Research on the algorithm of image processing and dimensional features: First,a series of image pre-processing operations are performed on the collected image,and the image is rotated and adjusted.Then perform edge operator processing on the image to get the edge contour and extract its corner pixels,and design an interest point screening algorithm based on the parameters required to calculate the dimensional features.Finally,based on the knowledge of mathematical geometry,the corresponding algorithm flow is designed for the relevant parameters of the screw dimensional features.3)The first stage of deep learning defect detection network research: First,introduced the advantages of deep learning in the field of defect detection,designed a cascaded deep defect detection network,and explained the structure of the network.Then use the supervised learning model to train the network and create relevant datasets.Finally,the method of transfer learning is used to fine-tune the pre-trained network model according to the tasks in this paper,and the network is trained using different Fine-tune methods.The final network model is obtained according to the optimal results of the experiment,and the first stage of the detection task is completed.4)The second stage of the deep learning defect detection network: First,it introduces and analyzes different methods and corresponding training modes for dividing the datasets in the network for defect feature detection,and proposes a new idea to construct the datasets according to the GAN network.Then according to this idea,create datasets for the task of this paper,and the process and advantages of the deep defect detection network model structure are analyzed.Finally,the experience of the first stage of network training is used to conduct semi-supervised training of the second stage of the network,and we analyze the experimental results of the defect detection of the network model.