
36
5 CONCLUSIONS
Content-based image retrieval applications usually
compute image feature vectors once and employ
them several times. Then, the reduction of feature
vectors is desirable in terms of computational time
and storage requirements. Since LDA and PLS are
able toidentify a feature vector as belonging to a par-
ticular class the performance they guarantee a bet-
ter performance. Nevertheless, PCA, ICA and RP
techniques also present an acceptable performance.
This work demonstrated the feasibility of reduc-
ing the dimensionality of descriptors, such as SIFT
and SURF, while maintain a similar accuracy or
even improve it when using less than ...