Robust Angle Invariant 1D Barcode Detection

Transcription

1 Robust Angle Invariant 1D Barcode Detection Alessandro Zamberletti A. Zamberletti, I. Gallo and S. Albertini Department of Theoretical and Applied Science (DiSTA) University of Insubria ACPR November 2013 Okinawa, Japan A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 1 / 15

2 Introduction Goals: detection of 1D barcodes from real world images fast execution on a low computational power device robust to lighting variations, background clutter, rotation and distortions A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 2 / 15

3 Related Works Existing 1D barcode detection methods: gradients analysis O. Gallo and R. Manduchi, Reading 1-D barcodes with mobile phones using deformable templates, IEEE PAMI, scan lines S. Wachenfeld, S. Terlunen and X. Jiang, Robust 1-D barcode recognition on camera phones and mobile product information display, Mobile Multimedia Processing, R. Adelman, M. Langheinric, and C. Floerkemeier, A toolkit for barcode recognition and resolving on camera phones, MEIS, canny edge detector E. Basaran, O. Ulucay and S. Erturk, Reading barcodes using digital cameras, IMS, Issues: slow rotation A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 3 / 15

4 Ideas Novel detection approaches: optical character recognition methods numbers are not part of the barcode standard Hough transform for lines detection successfully used for reading 1D barcodes R. Muniz, L. Junco and A. Otero, A robust software barcode reader using the Hough Transform, IISA, A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 4 / 15

5 Hough Transform Detection of 1D barcode patterns in the Hough Transform space using a simple MLP network: pros cons rotation invariant fast sensitive to noise requires training A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 5 / 15

6 Supervised Angle Detection Canny Edge Detector MLP Network n m A H A N c(i, j ) ct (i, j ) θ θ ρ ρ hl θ r Hough Transform 1 a regular grid n m of cells c 1,..., c n m is superimposed over the Hough accumulator matrix A H 2 c is given as input to the MLP 3 the MLP produces a new cell c t c t(i, j) = { 1 if c(i, j) denotes a barcode bar in I. 0 otherwise. 4 the processed cells are combined together to generate A N A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 6 / 15

7 Supervised Angle Detection A N θ r h l 5 given a row r A N, the sum of its elements denotes the likelihood l r of a barcode appearing in I rotated by θ r 6 we build the histogram of likelihoods h l 7 rows associated with max elements of h l denote barcodes A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 7 / 15

8 Supervised Angle Detection - Details MLP: 3 layers size of each layer: n m trained using rprop (parameters from Igel et al. 2005) 200 training patterns (150 bgs, 50 fgs) Training patterns - c i A H, (in ci, out ci ): in ci - vector representation of c i out ci - elements of in ci denoting barcode bars are assigned 1, the others are assigned 0 A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 8 / 15

9 Supervised Angle Detection - Datasets Dataset WWU Muenster Barcode DB* ArTe-Lab 1D Medium Barcode** Rotated Barcode DB** No. of Imgs Device Barcode Rotation Nokia N95 Nokia 5800 Various ±30 // ±180 All the datasets are available online for download: * ** A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 9 / 15

10 Supervised Angle Detection - Evaluation Results: ArTe-Lab 1D Medium Barcode 100% OA θ WWU Muenster Barcode DB & Rotated Barcode DB 95.5% OA θ Time: 200 ms per image Metric: OA θ tp = tp+fn+fp cell height cell height cell height WWU Muenster Barcode Database cell width ArTe-Lab 1D Medium Barcode Dataset cell width Rotated Barcode Database overall angle detection accuracy overall angle detection accuracy overall angle detection accuracy cell width A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 10 / 15

11 Bounding Box Detection Detected Barcodes A N h c S b θr θ ρ hl h r S b I Sb 1 rotate I by θ r (detected during the previous phase) 2 get the segments parallel to the vertical, as in Matasyz et al., obtain a binary image I Sb 4 define two histograms h r S b and h c S b describing the intensity profiles of the rows and columns of I Sb respectively A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 11 / 15

12 Bounding Box Detection Detected Barcodes h c S b h r S b I Sb 5 smooth the histograms to remove low value bins corresponding to isolated non-barcode segments 6 the detected bounding boxes correspond to the intersection area between the rows and the columns associated with the non-zero bins of h r S b and h c S b A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 12 / 15

13 Bounding Box Detection - Evaluation Time: 70 ms per image Dataset OA bb WWU Muenster Barcode DB 0.83 ArTe-Lab 1D Medium Barcode 0.86 Rotated Barcode DB 0.84 Metric: OA bb tp = tp+fn+fp tp = # of barcode bounding boxes correctly detected a bounding box d b is correctly detected iff bb b d b 0.5 bb b d b A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 13 / 15

14 ZXing and Demo Dataset Reading Accuracy ZXing Our ZXing acc WWU Muenster Barcode DB ArTe-Lab 1D Medium Barcode Rotated Barcode DB A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 14 / 15

15 Q&A Thank You! Department of Theoretical and Applied Science (DiSTA) University of Insubria Varese, Italy web: A. Zamberletti et al. - Robust Angle Invariant 1D Barcode Detection - ACPR13, Okinawa, Japan 15 / 15