A Defect Detection Method Of Metal Gear Machined End Surface Based On Salient Region Extraction And Modified YOLO-V3

Defect detection of metal gear Machined end surface is an important process in gear manufacturing.In precision machinery and equipment,defects such as scratches,dents,bumping and etc.,may lead to many potential problems,including deterioration of working conditions,abnormal noise,and even machinery malfunctioning,resulting in huge losses.Metal gear Machined end surface defects are usually caused by residual iron filings,improper clamping,physical collision,etc.,and they are prone to batch problems.At present,the manufacturing enterprises mainly use manual sampling to detect the surface defects of gears.There are problems such as relying on manual experience,misdetection and missed detection,and long time-consuming.At present,in the industrial field,many defect detection methods based on machine learning and deep learning have been carried out.However,the surface accuracy of metal gear is high,and the defect detection accuracy is required to be higher.There are some troubles in existing methods,such as omission of inspection,false inspection,reduced efficiency and instability of the defect inspection process.In this paper,the object detection methods and visual salient methods are introduced into the detection of gear surface defects,a defect detection method of metal gear Machined end surface based on salient region extraction and modified YOLO-V3 is firstly proposed to carry out accurate and fast defect detection and localization.This method is accurate,fast and low cost,which can be applied to replace the manual detection.Using this method can realize rapid automatic detection,avoiding batch process problems in time and reducing gear scrap loss.In this paper,the defects,such as scratches,dents,bumping and etc.,are usually very small and similar with burrs,oil stains and non-machined areas,which makes their figures not obvious,not prominent,and hard to recognize.Existing defect detection methods based on machine vision have some troubles,such as lack of detection,error positives,low efficiency,and indeterminacy and so on.Existing image processing-based defect detection methods have problems such as missed detection,false detection,low efficiency and instability.This paper proposed a method of metal gear Machined end surface based on salient region extraction and modified YOLO-V3.The main research includes:(1)Propose a metal gear Machined end surface defect salient region extraction method(MGSRE).Considering the specular reflection characteristics of the machined area and the non-machined area of the metal gear,the L channel in the Lab color space is very sensitive to the brightness.Then the frequency domain tuning saliency algorithm(FT)is used to generate the saliency map of metal gear surface image.After Gaussian filtering,the image is transferred to lab color space to generate the mask map of gear machining surface area.At last,through the logic AND operation of gear image and mask image,the area of gear machining is extracted to eliminate the influence of interference factors.(2)Propose an improved YOLO-V3(Res Net YOLO).Adopt the Res Net structure,a high-resolution lightweight feature extraction backbone network(Res Net-21)is designed.Firstly,Resnet-21 adopts 16 x down-sampling,which can better extract the features of small defects,and uses Leaky Re Lu as the activation function of each convolution module to enhance the stability of network training process.Secondly,considering that the low-level features are usually lost in the process of down-sampling by pooling operation,cancel the pooling layer,and each stage uses convolution kernel with step size of 2 to down-sample.Finally,by reducing the number of channels and introducing the Bottel Neck structure with a 1 * 1 convolution kernel,the excess computational problem caused by several additional parameters owing to the increase in the functional map is resolved.The Res Net-21 achieves efficient extraction of high-resolution features.(3)Propose a multi-scale fusion method that fuses the minimum-scale feature map output by Res Net-21 with the medium-scale and the maximum-scale.The maximum(152 * 152)scale,medium(76 * 76)scale and minimum(38 * 38)scale are used as defect detection features.Local fusion of features is realized by stacking features of different scales and combining with convolution kernel.This method makes the feature map have the characteristics of high resolution and high semantic information.At last,the three-scale feature maps are classified and located by the multi-classifier module.(4)The trained metal gear surface defect detection model is tested and verified in a gear manufacturing workshop,and the experimental results are analyzed.The results show that,both the m AP and mean Recall of the metal gear defect detection method are97% and one single detection time is 70 ms,which shows the effectiveness of the proposed method.