Liu Dichao

Abstract

Fine-grained vehicle recognition (FGVR) is an essential fundamentaltechnology for intelligent transportation systems, but very difficult becauseof its inherent intra-class variation. Most previous FGVR studies only focus onthe intra-class variation caused by different shooting angles, positions, etc.,while the intra-class variation caused by image noise has received littleattention. This paper proposes a progressive multi-task anti-noise learning(PMAL) framework and a progressive multi-task distilling (PMD) framework tosolve the intra-class variation problem in FGVR due to image noise. The PMALframework achieves high recognition accuracy by treating image denoising as anadditional task in image recognition and progressively forcing a model to learnnoise invariance. The PMD framework transfers the knowledge of the PMAL-trainedmodel into the original backbone network, which produces a model with about thesame recognition accuracy as the PMAL-trained model, but without any additionaloverheads over the original backbone network. Combining the two frameworks, weobtain models that significantly exceed previous state-of-the-art methods inrecognition accuracy on two widely-used, standard FGVR datasets, namelyStanford Cars, and CompCars, as well as three additional surveillanceimage-based vehicle-type classification datasets, namely Beijing Institute ofTechnology (BIT)-Vehicle, Vehicle Type Image Data 2 (VTID2), and Vehicle ImagesDataset for Make Model Recognition (VIDMMR), without any additional overheadsover the original backbone networks. The source code is available athttps://github.com/Dichao-Liu/Anti-noise_FGVR