Research On Image Detection Methods And Key Technology Based On Subpixel

With the rapid development of the precision manufacturing industry, the demanding for image detection is increasing. At the same time, as the camera resolution raises and the image processing algorithms get maturer, the computer vision technology is widely used in daily detection of production line. Motivated by this real requirement and existing problems, this thesis aims to study and implement an image detection system based on subpixel technology.To quickly eliminate some obvious abnormal products, the moment invariants broadly used in the image retrieval field is introduced to the image detection field. On the basis of polar-radius-invariant-moment, we construct a new polar-radius-moment and propose a combined-moment invariants method, which combines the first four lower moments of Chen moment invariants and the first three lower moments of the new polar-radius-moment. This method is applied to eliminate obvious abnormal products quickly. As a result, the average time cost of every product test is reduced and the efficiency of the detection system is improved.The Common phenomenon exists that some commercial detection system has low precision, but high precision algorithms are always infeasible to be applied to industrial production line due to its slow calculation. Inspired from this point, we present a subpixel partioning algorithm based on multiquadirc radial basis interpolating function. This approach enables accurate positioning of subpixel level edge and efficiently improves the detection precison of the system. On the other hand, traditional subpixel interpolating methods do not fully employ the gray-level distribution and variation information along the vertical direction. To overcome this limitation, this novel method constructs a multiquadirc radial basis interpolating function, regards the gradient as the function of horizontal and vertical coordinates, and conducts interpolations on the subpixels of curved surfaces. To avoid large calculation on curved surface interpolation, we attempt to improve the interpolation algorithm and take advantage of the grey-level distribution and variations along the vertical direction. Then, we project the obtained interpolation function onto the horizontal direction to do curve interpolation. In this way, our method can greatly reduce the computation complexity so as to satisfy the time requirement for realtime online detection of industrial production line.To further enhance the algorithmic efficiency, with regard to concret algorithms involving a great amount of similar calculations and comparison operations, we apply the heap structure rather than the classical two-dimensional array. This is able to avoid comparing many function values to attain the maximum, which significantly reduces the time complexity. In addition, we have implemented the thread pool under LINUX and decomposed the subpixel algorithm, thus achieving the parallel running of multi-thread sequences on the algorithmic level and high efficiency of the detection system.In summary, this thesis designes and implements the image detection system under LINUX. We concentrat our focus on how to increase the detection accuracy of the algorithms and how to efficiently reduce the time complexity. Experimental results have demonstrated the advantages of the proposed approaches.