A visually salient approach to recognize vehicles based on hierarchical architecture

In order to recognize multi-class vehicles, traditional methods are typically based on license plates and frontal images of vehicles. These methods rely heavily on specific datasets and thus are not applicable in real-world tasks. In this paper, we propose a novel method based on a hierarchical model, HMAX, which simulates visual architecture of primates for object recognition. It can extract features of shift-invariance and scale-invariance by Gabor filtering, template matching, and max pooling. In particular, we adopt a model of saliency-based visual attention to detect salient patches for template matching, also we drop inefficient features via an all-pairs linear SVM. During experiments, high accuracy and great efficiency are achieved on a dataset which has 31 types and over 1400 vehicle images with varying scales, orientations, and colors. With comparisons with Original-HMAX, Salient-HMAX, and Sifted-HMAX model, our method achieves classifying accuracy at 92% and time for each image at around 1.5s, while reduces 73% of the time consumed by original HMAX model.