Research On Application Of Improved YOLOv3 Algorithm In Defect Detection Of Burr Cylinder Liner

Burr cylinder liner is an important component of automobile engines,and its appearance quality seriously affects the service life and fuel consumption of automobile engines,so it is very important to check its appearance quality.The burr cylinder liner has many kinds of appearance defects and different judgment standards,which makes the detection work more difficult.At present,the appearance defect detection of burr cylinder liners is generally inspected manually.Such a detection method has the disadvantages of low efficiency,low accuracy,and large workload.Current detection methods based on image processing can efficiently identify appearance defects,but because the color of the burr cylinder body and the defect area are similar,it is difficult to obtain high-quality pictures that contain multiple defects,which limits the use of image processing algorithms.The article combines deep learning target detection and image processing algorithms to detect five kinds of defects in the appearance defects of the burr cylinder liner,overcoming the limitations of traditional image processing and manual detection.In the experiment,the qualitative analysis and quantitative analysis of the defects are carried out respectively,so that the detection algorithm is more accurate.After comparative analysis,the efficient YOLOv3 algorithm is selected.In order to better apply the defect detection of the glitch cylinder liner,its feature extraction network is improved.In the experiment,a dense connection mechanism is used to reduce the loss of information transmission between feature layers,alleviate the problem of gradient disappearance caused by deep networks,and speed up the update of the network;in order to improve the recognition accuracy of the algorithm,feature re-use is used for each convolutional feature Calibration algorithm makes the network pay more attention to the channels that have an effect on defect identification;the difference in defect size is an important reason for the accuracy of the algorithm.In the experiment,it is proposed to use deformable convolution instead of standard square convolution to increase the network’s ability to detect objects of different sizes.Finally,for the analysis of the detection network,the NMS algorithm used in the source algorithm has a problem of missing detection for defect detection,and a soft-NMS algorithm with linear weighting is proposed to replace the NMS algorithm.After experimental comparison,the soft-NMS algorithm reduces the missed detection rate of the defect data set.In order to distinguish the YOLOv3 algorithm,the improved YOLOv3 network is defined as Advance-YOLOv3.In verifying the effect of the improved algorithm,comparative experiments were conducted from classification and detection.In the experiment,keep the same data set,experimental platform,and experimental parameters.Compared with the two classic classification networks VGGNet-19 and Res Net-152,the highest classification accuracy is 83.7%,which is an increase of 7.5% and 1.1%,respectively.In the detection task,the YOLOv3 source is compared respectively The model and Faster R-CNN network have a mean accuracy of 79.6%,which increases by 3.5% and 4.4% in the detection accuracy index,and performs poorly in the detection time index.The average time for a single image is 86 ms,which is higher than the two algorithms 4ms and 4ms.10 ms,the main reason for the analysis is that the deformable convolution learns the deformation parameters;it is better than the other two algorithms in the missed detection rate index,reducing 7.5% and 3.3% respectively.The comparison experiment of classification and detection proves that the improved YOLOv3 algorithm has improved detection accuracy and missed detection rate,indicating the effectiveness of the algorithm.