Research On Feather Segmentation And Quality Analysis Based On Machine Vision

As a big producer of badminton,China accounts for 90%of world's badminton production.So far,some important feather sorting process still needs manual operation,which is not only slow and has low accuracy but also do harm to workers'health,then automatic feather sorting system came into being.An efficient sorting system will bring huge benefits to industries.However,due to the complexity of feather classification criteria and interference of the production environment,automatic segmentation of feather images and inspection of feather quality are still problems that need to be solved in feather sorting system.This thesis proposes a superpixel-based segmentation and defect detection method.By replacing common image partition methods like grid and sliding window,superpixel-based segmentation method uses local gray value and textural information to locate the region containing feather shafts.This segmentation method also avoid the effects of uneven lights.Based on the location result,we propose an improved Otsu algorithm to segment feather shaft and feather leaf within the local super pixel range and obtain a better result than original Otsu method.On the basis of feather segmentation,we combine the gabor filter and different statistical texture features to detect feather defects.According to the defect recognition rate of using different features alone,we finally fuse classifiers based on rotation invariant uniform LBP and HOG and obtain better recognition results.Defect detection results can be used to identify low-grade feathers.In addition,we select appropriate texture descriptors to extract feather's features and train classifiers to grade feathers.We apply deep neural network to feather grading.To the best of our knowledge,we are the first to use CNN to solve problems of such kind.In this thesis,we analyzed the effects of network parameters such as the number of layers and the number of kernels on feather grading results.Besides,we also introduce global average pooling and multiscale feature into the model.These all have reference value to the construction of convolutional neural network for feather grading.Experiments show that the grading accuracy based on CNN is much better than the one based on traditional method.Finally,we implement an automatic feather sorting system,which combines the feather segmentation,defect detection methods and classification methods.The whole process of feather image acquisition,feather classification and many other detection and feathers delivery is realized by the sorting system combined with hardware pipeline.The feather grading module,as the core part of the system,makes up for the blank of automatic feather grading technology in badminton production process.The system has been put into the actual production environment.From the trial results,the system is innovative and competitive.We perform several sets of contrast experiments on feather images collected in industrial production to explore the effects of machine vision related techniques on feather sorting.Our feather segmentation method is proved effective in different conditions and our classification method is also valuable for practical application.