Boosted Particle Swarm Optimization of Gabor Filter Feature Vector

Many feature vectors used today for classification of images require a significant amount of computation. These feature vectors are high dimensional. An attempt to build an optimized feature vector using Gabor filters is discussed. AdaBoost and Particle Swarm Optimization techniques are used to select Gabor filters which maximize the distinction between positive and negative samples, when comparing the phase at a single location in the image. The feature vector is therefore optimized for a particular classification task, and its performance in pedestrian classification is compared to the HOG feature vector. Although the results do not currently match existing feature descriptors, this method still shows promise due to its extremely low dimensionality, and should drastically improve when further dimensions are added.