Leather Defects Recognition And Classification Based On A Parameter Optimized Residual Network

With the improvement of people’s living standards,the consumer demand for leather products has also changed from practical to quality.The demand of consumers stimulates the manufacturing field to continuously improve its production technology to ensure the high quality of leather products.Among them,there may be many kinds of defects on the leather surface of leather raw materials in the process of processing.It is helpful to improve the quality of leather products by accurately identifying the defective leather materials before the production of leather products.At the same time,the leather with defects can be treated differently according to the types of defects,so the automatic classification of different defects can further improve the subsequent work efficiency.The detection and classification of leather surface defects are mainly done manual in China.This manual method is influenced by subjective factors,and workers are prone to misdetection and misjudgment due to external factors such as emotions and working hours.At the same time,the difference in the efficiency of manual operations has seriously affected the progress of the entire processing process.Therefore,the automatic classification of leather surface defects is of great significance to the improvement of the overall efficiency of the leather manufacturing industry and the reduction of production costs.At present,the domestic and foreign researchers have proposed many different detection algorithms for the detection of leather surface defects,but these algorithms are for the detection of leather raw materials for defects,and cannot automatically classify various types of defects.Based on the parameter optimization residual network,we study and analyze the automatic classification of defective leather.The main contributions of this thesis are summarized as follows:1.To solve the problem that the current researches conducted on the leather defect classification failed to classify the types of the leather defects automatically and objectively,a framework of the classification of the leather surface defects based on a parameter optimized residual network is proposed.In this thesis,the defects of the leathers will be automatically divided into five categories: the scratches,the rotten surfaces,the holes,theneedle eyes and no defects.The classification results are able to effectively improve the subsequent leather processing quality.2.The least squares method is used to fit the computer numerical simulation data to determine the appropriate size of a sliding patch window.The defects of the leather are inside the regions with different sizes on the leather sample images.Thus,how to determine the appropriate size of the sliding patch window is a challenge issue.A too small sliding patch window cannot fully contain the defect area,which will make the corresponding defect feature extraction inaccurate in the network training.On the contrary,a too large sliding patch window may weaken the effectiveness of the extracted defect features.In order to solve this problem,the comparative computer numerical simulations of different sizes of the sliding patch windows are conducted.Based on the computer numerical simulation results,the final size of the sliding patch window is determined by the least squares method.3.The appropriate size of the data volume is verified by the computer numerical simulations.In the deep learning network,the small training data sets may result in overfitting.On the contrary,the large scale training data sets can improve the network training precision,but it will greatly increase the computer numerical simulation workload.Therefore,the appropriate size of the data set is determined by examining the computer numerical simulation results on different scales of data sets.4.Through experimental analysis,the parameters of the deep residual network are optimized,and the performance of different networks is verified by comparative experiments using the same leather defect data set.The experimental results show that the parameter-optimized residual network used in this thesis effectively detects and classifies leather defects,and outperforms common network models.The classification accuracy reaches 94.6%...