Research On The Method Of Identifying Surface Defects In Car Seat Leather

The defects on the leather surface,including insect holes,neck lines,cracks,holes,scars,rotten pores,creases,and foreign objects,seriously affect the quality of the final product.At present,in the production process of automotive leather seats in China,manual methods are still used for visual inspection of leather surface defects,which is inefficient and costly.The inspection standards rely on personal subjective experience,and there is a significant risk of misjudgment and missed inspection of quality.This article takes the leather defect material of a certain car seat as the research object,designs a leather defect detection algorithm based on ResNet,and proposes a histogram based on pixel spatial distribution features for multi threshold segmentation of leather surface defects.Compared with traditional gray histogram based multi threshold segmentation,the effect is significantly improved.The research content of this article is as follows:(1)This study first collected common leather defects in actual production,and constructed a leather defect dataset consisting of 8 types of defects such as wormholes and cracks,and 9 categories of a normal good product.The original data set was expanded by translation,rotation,and multi-scale scaling,respectively,and the classification model was constructed by ResNet50 that had been pre trained by Image Net data set.Compared with the effect of data amplification,the overfitting effect of model training was significantly improved by using the data amplification method in this paper,and the accuracy of classification was increased from 75.6% to 98.6%.Then,by comparing the classification performance of ResNet50,ResNet18,ResNet34,and Google Net models for 8 types of defects,the results showed that the ResNet50 model had an accuracy of over 98% on the leather surface defect dataset,a recall rate of over 96% for each defect,and an F1 value of over 96%,making it the optimal model.(2)In response to the shortcomings of traditional grayscale histograms that only contain information on the number of pixels at each grayscale level and lack spatial distribution information at the grayscale level,this paper proposes a spatial distribution feature histogram constructed based on the spatial distribution features of grayscale pixels.The histogram is combined with PSO algorithm,Otsu method and minimum cross entropy method to segment leather defect samples with multiple thresholds.Experiments have shown that the segmentation effect of spatial distribution feature histograms is significantly improved compared to the image grayscale histogram method.(3)In order to study the characteristics of the spatial distribution feature histogram proposed in this article,three modules that constitute the spatial distribution histogram were studied.Based on these three modules,histograms were constructed for multi threshold segmentation,and the effectiveness of each module was analyzed through ablation experiments.The experiment shows that the segmented PSNR value will gradually increase with the integration of modules,which can confirm the effectiveness of the three modules of the spatial distribution histogram.At the same time,this paper studies the hyperparameter of the graph structure complexity coefficient in the spatial distribution feature histogram: the weight c,through the design gradient of the weight c value,this paper compares and analyzes the impact on the image multi threshold segmentation results when the weight c value is different,and also obtains the optimal value of the weight c in the leather defect scene.