Object Recognition Using Deep Neural Network with Distinctive Features
Hyun Chul Song, Farhan Akram, Kwang Nam Choi · 2018
In this paper, a new object recognition method using statistically weighting Multi-Layer Perceptron (MLP) is proposed. It uses visual distinctive features, which are computed using Bag of Visual Words (BoVW) framework. The proposed method has the following three main steps. At first it represents the images into their respective co-occurrence matrices, which are vectorized using BoVW and gives distinctive features. Then it computes weights from the histograms of visual words for each class. Finally, the statistically weighting distinctive features are applied to the testing image set to find the object class. In the proposed method, we improved MLP by introducing the weighted visual words, which are extracted by sampling the patches from the current image. From the Caltech 256 dataset, four classes namely pedestrians, cars, motorbikes and airplanes are used for the classification accuracy comparison between the MLP based artificial neural network (ANN) and the proposed method. The experimental results show that our method outperforms traditional MLP yielding an average classification accuracy of 89.60%, which is approximately 6.3% more than the compared MLP.