Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 98

Fig. 3.19 Average Precision Rate, APR (%) obtained by the adaptive cosine network through
pseudo-RF, using 25 video shots for queries. For the pseudo-RF result, the signal propagation was
allowed to flow forward and backward for up to 20 iterations, and only the result after 1, 3, and 20
iterations are shown, for comparison with the non-adaptive method that used the cosine metric as
the similarity matching function

affected the results and
confirmed the reports of other studies [ 103 , 107 , 323 ] with regard to the value for

It was observed that the values for ʾ , ʱ and ʲ

.
However, the identification of the proper values for these parameters was completed
conveniently as they were usually found in certain ranges. It was also observed
that without applying the threshold level

ʱ

, only modest improvement is initially
obtained, and all the nodes became increasingly activated. This led to a longer
processing time and to a random ordering of the videos.

ʾ

3.6 Summary

Automation is critical for enhancing learning efficiency and/or improving retrieval
performance. The automation is done through a pseudo-relevance feedback that
iteratively re-ranks database entities, in both fully-automatic and semi-automatic
modes. This chapter presents various techniques for pseudo-relevance feedback,
including dynamic self-organization methods and the adaptive cosine network. Both
compressed and uncompressed image databases, as well as video applications are
covered.