Research On CUDA Based Parallel Image Processing Model And Algorithms For Fabric Defect Detection

Fabric defect detection is one of most significant procedure effecting the manufacturing efficiency and quality control in textile industry. The manual detection is time-consuming, labor-intensive and devoid of consistency and reliability due to many subjective factors. Computer vision technology has been applied widely in industrial surface detection with the development of computer hardwares and digital image processing techniques. Now fabric defect detection systems on the market primarily come from abroad. It can only detect obvious defect types but the cost is high. A large number but small size is the major characteristic of China’s textile enterprises, so the high cost systems from abroad can’t be accepted by most enterprises and the manual detection is the main method in China. In recent years, there appeared many defect detection algorithms and new systems at home, but subjected to the complexity and the cost of these systems they have not been put into application.GPU (Graphics Processing Units) has more wide application in engineering. It introduces new ideas and solutions for improving computational efficiency. Based on these, a set of fabric defect detection algorithms suitable for parallel processing are proposed and optimized by CUDA (Compute Unified Device Architecture) in this paper. The automatic fabric defect detection system based on CUDA is discussed with common computers as a platform. The research offers the theory and practice foundation for structuring high quality detection system with competitive price and promotes the development of fabric defect detection technology and application in China.GPU is introduced into fabric defect detection system as the main processor and a set of new detection algorithms and optimized strategies based on CUDA are proposed in this paper. The main works and contributions are as follows:· To solve the application of CUDA technology in digital image processing, a CUDA-based parallel processing model is proposed. This model introduces basic procedures of digital image processing and basic optimization methods in CUDA platform. Although image processing algorithms are different, there is a common model when applying CUDA. According to this model, several common algorithms in digital image processing are parallelized based on CUDA and experimental results show the high efficiency of the model.· A new fabric defect judgement algorithm based on direction operators and its optimization strategy using CUDA are proposed. The detected image is separated into smaller windows and a characteristic value is extracted from each window based on the information of the gray value and orientation. By accumulating the characteristic values in rows or columns, the difference between the normal texture and the defect areas is increased and this makes the judgement result more accurate and practical. Several parallel strategis are designed to optimize the proposed judgement algorithm and proved to greatly improve computational efficiency by the experimental results.· A new fabric defect segmentation algorithm based on regional growing PCNN (Pulse-Coupled Neural Networks) and its optimization strategy using CUDA are proposed. The improved PCNN model and the regional growing theory are combined together. The pixels in the detected image are mapped to the neurons in the PCNN network and the grayscale statistic information of the defect-free image is introduced into this model. Neural network algorithm itself has the immanent parallel mechanism, so it has the possibility of parallel compute on a large scale. The proposed segmentation algrithem can be greatly accelerated by the optimization strategy based on CUDA.· A fabric defect classification method using Support Vector Machine (SVM) and Gray-Level Co-occurrence Matrix (GLCM) is put forward and carried out. Compared to the judgement and segmentation algorithms, the classification method has non-strict demand for real-time. The defect blocks are extracted from the fabric image based on preceding segmentation results. The characteristic values are obtained by computing the GLCM of each block and input to the SVM classifier. Classification results indicate the high recognition rate of the method.