Leber Recognition And Its Parallel Speedup On GPU

In the underwear factory, it should be ensured that the goods in the same box have the same information, such as size and number etc. All the critical information is printed on the label, which is attached to the underwear. For a long time, Quality Inspectors check the consistency with their naked eyes. But this does great harm to eyes. Meanwhile, it results in high probability of errors. Thus, an automatic underwear detect system is consumingly demanded. The core of the system is character recognition on the label.The most widely application field of Character recognition is license recognition. Differing from license recognition, label recognition has its own difficulties, for example, characters in labels contort more easily, their location is more complex and their length is un-fixed. These difficulties determine that methods in license recognition are no longer suitable to label recognition. We need new ones.The main modules of the system are:data collection, pretreatment, label location, strings separation, OCR recognition and fault tolerance.1. In the data collection module, images are collected by Microview MVC-II-3M USB camera. In company with the camera's SDK, images are saved to the specified path in order to process them later.2. In pretreatment module, the collected digital images are firstly transformed to gray ones. After gray conversion, noise filtering and binarization, they come to be binary images with less noise and more contrast.3. In label location module, edge detection is done to the binary images first. If the original image is collected in the ideal case, then the next step is to scan the edge of the label to get its three boundary lines, and the location is finished. Otherwise, the three boundary lines can be got after doing Hough transformation and clustering. Then the label will be located accurately. According to the location result, the label image can be gained by cropping.4. In strings separation module, two strings are cropped according to their distribution on the label. After doing morphological operation to the horizontal string and rotation to the vertical one, combine them into one image by way of next module.5. In OCR recognition module, do OCR operation to the results of the previous steps with Microsoft Office tools-MODI.6. Because the OCR recognition's accuracy rate is not 100%, the system adds fault tolerance module to improve the whole accuracy, which is done under the characters' practical significance.In label location module, Hough transformation is used. It occupies a considerable percentage time of the whole, so its parallel speedup version on GPU is promoted. Because of Hough transformation' large amount of calculation, the parallel algorithm is completed by a series of optimization such as thread mapping, data accesses, data reuse etc., based on a unified GPU computing architecture. These optimizations include:1) pixel to thread correspondence base on pixels' independent; 2) dimension reduction storing to achieve coalesced accesses; 3) reading data in shared memory. Compared with serial implementation on CPU, the parallel version on GPU outperforms both in bandwidth and GFLOPS. In addition, the total performance of the parallel version could be 10 timeshare greater than that of the CPU counterpart.