LightWeight Deep Convolutional Neural Networks for Vehicle Re-Identification using Diffusion-based Image Masking

Re-Identification of Vehicles (Vehicle Re-ID) in surveillance videos is an important and emerging topic in computer vision and intelligent systems. In Vehicle Re-ID, a specific vehicle is searched through non-overlapping camera viewpoints, which is a difficult task due to varying image conditions (e.g. scale, variations of appearance, illumination). For vehicle Re-ID, low level features such as shape and color can be utilized. Recently, deep learning-based vehicle Re-ID solutions became popular. Many of these deep learning solutions use pretrained large convolutional neural networks (CNNs) to perform vehicle Re-ID. Most of the pre-trained networks for object recognition and retrieval work with a large input image size (e.g. 224 x 224 x 3) and many hidden layers/parameters. In this work, we propose a lightweight deep CNN model together with diffusion-based image masking. The lightweight network trains input images of size 32 x 32 x 3, it also has a smaller number of parameters to train in comparison to large pre-trained networks. We use three lightweight CNN networks, the first network trains original colour vehicle images, the second network trains vehicle images after masking with fully diffused shape information image, and the third network trains vehicle images after masking with partly diffused shape information image. Feature vectors extracted from these three networks are used to generate three similarity scores. Then, these three similarity scores are fused to produce an overall similarity score for retrieval. We demonstrate that the proposed work performs better than many large pretrained networks on VeRi-776 dataset.