Iterative tensor tracking using GPU for textile fabric defect detection

This paper presents an efficient real-time implementation of an unsupervised textile fabric defect detection algorithm called ITT using the concept of iterative tensor tracking on graphics processing unit (GPU). The algorithm adopts a new local image descriptor, Spatial Histograms of Oriented Gradients (S-HOG), which is shift-invariant, light insensitive and space scalable. For a given textile fabric image, ITT iteratively updates and then analyzes S-HOG using tensor operations, in particular tensor decomposition to detect textile defects. To speedup the calculation required, ITT is implemented on the GPU using the Compute Unified Device Architecture (CUDA) programming model. The respective computational efficiencies of implementing ITT on the GPU and on the CPU are compared by using experiments. The results demonstrate that the computation speed of the former is on average thirty times and ten times faster than that of the later for updating the S-HOG and for detecting defects respectively because of its parallel processing nature.