Logo Recognition

9

Conclusion

CONTENTS

9.1 Book summary ............................................................ 130

9.2 Contribution .............................................................. 130

9.3 Future work ............................................................... 131

9.4 Book conclusion ........................................................... 132

Logo recognition is of great interest in the document and shape matching do-

main. Logos can act as a valuable means to identifying sources of documents.

However, logos are 2D shapes of varying complexity, with interior and exte-

rior contours that are not necessary connected. Hence the recognition process

seems to be difficult because of its complexity. Researchers have investigated

the problem of logo recognition [46, 189, 91, 234, 213, 36, 143]. Although some

very effective results have been found for clean logos, they can hardly be ro-

bust with corrupted logos, such as strips obstructing the logo in unpredictable

positions. Other research on logo recognition by neural networks [26, 25, 73]

can deal with noisy logos; however, it requires a lot of training before it can

be used.

In order to provide better distinctive capability for recognizing logos under

adverse conditions such as different scale/orientation, broken lines and added

noise and occlusion, this book proposes a new logo recognition approach, i.e.,

a hybrid of structural feature based and template based techniques, which

includes two aspects:

• Consistent logo representation: Extract local line pattern features which

are invariant to scale, orientation, translation and reasonable amounts of

noise and occlusion.

• Effective logo recognition: Find a suitable similarity/dissimilarity mea-

sure which is efficient to compute, tolerates reasonable amounts of noise

and occlusion and degrades gracefully as these tolerances are exceeded.

129

130 Logo Recognition: Theory and Practice

9.1 Book summary

This book has investigated an approach (combining the structural and tem-

plate matching techniques) that can represent and recognize complex shapes

(i.e., logos) under noisy conditions. It can tolerate reasonable amounts of noise

and occlusion, and degrade gracefully as these tolerances are exceeded. On the

other hand, since the method is based on line segments, it is easy to imple-

ment and demands less storage space. This work has involved the following

investigations:

• Polygonal approximation: Transform raw logo images into consistent Line

Segment Maps (LSM).

• Indexing: Investigate and search for effective line pattern features that

can be used to index the database to generate a moderate number of

likely models with respect to a test image.

• Matching: Propose an improved Line Segment Hausdorff Distance (LHD)

measure to screen further and generate the best matches.

In this book, numerous improvements have been made, including a con-

sistent feature point detection method, a robust normalization process, an

effective indexing approach and a reliable MLHD algorithm that is the most

important part in the proposed system. An in-depth study of the proposed

technique has also been carried out on logos. The test images come from seven

sources, i.e., regenerated, strip corrupted, partially occluded, mixed noise (i.e.,

spot and white Gaussian noise) corrupted, cylinder projected, skewed and for-

eign logos. Encouraging results have been observed. These results show that

the proposed technique is invariant to scale, orientation, and translation and

tolerates reasonable amounts of noise and occlusion. It degrades gracefully as

these tolerances are exceeded. Compared with other works on logo recognition,

no existing technique can tolerate as many distortion types as that proposed

in this book. On the other hand, the proposed method will eventually fail

when severe distortion occurs, or when the normalization process fails to find

a single reference line. Nonetheless, to the best of our knowledge, other ap-

proaches cannot tackle such distorted images as well. The proposed approach

can work reasonably well as long as one suitable reference line can be found,

while other approaches are not likely to succeed.

9.2 Contribution

This book has made the following contributions: