An Effective Tread Pattern Image Classification Algorithm based on Transfer Learning

Tread pattern image classification is an important means in providing useful clues in traffic accident control and crime case solving. This paper proposes an effective tread pattern classification algorithm with feature fusion based on transfer learning. The algorithm consists of four parts, (1) Transfer the knowledge of a pre-trained CNN model on ImageNet dataset to produce a new model for the task of tread pattern classification by fine-tuning the model parameters using tread pattern image data. (2) Extract features from multiple fully-connected layers as high-level features of tread pattern images. The concept of transfer learning solves the problem of lacking large training dataset. (3) Histogram of Oriented Gradient (HOG) is calculated as the low-level feature of the tread pattern images. (4) The features from the CNN model are combined with HOG as fusion feature, which is used to train SVM classifier for tread pattern image classification. Experimental results demonstrated the outstanding performance of the proposed algorithm for tread pattern image classification.