Research On Vehicle Type Recognition In Surveillance Images In Few-Shot Learning

Vehicle type recognition(VTR)from surveillance data,which is the key module of intelligent transportation,received increasing attention in computer vision field and intelligent transportation.Deep learning methods such as convolutional neural networks(CNNs)has made great progress in vehicle identification for surveillance video due to its powerful representation performance.However,the key assumption of traditional deep learning is that the training data and the test data are independent and identically distributed(I.I.D),which requires sufficient vehicle data from the same surveillance system to be labeled to achieve deep learning methods in practical application.Considering that in the practical application,the surveillance system has different acquisition processes,it is very difficult to manually annotate the vehicle data.It is very difficult to obtain a universal and complete surveillance vehicle data set.These problems have largely limited the practical application of deep learning methods in surveillance vehicle type recognition.Based on the characteristics of vehicle type recognition from surveillance data,we focus on the few-shot learning method to solve the problem.By introducing transfer learning and active learning into deep learning,it not only retains the powerful feature extraction ability of deep learning itself,but also reduces the need for large-scale annotation samples.We proposed a surveillance-based VTR method based on deep learning and transfer learning.Our proposed method using only labels from Web data,which are easy to acquire from both Internet sources and vehicle manufacturers,to achieve the cross-domain recognition in surveillance vehicle data.In order to overcome the gap in the distribution of vehicles type between the Web data and the surveillance data,we introduced a transfer regularization term to the objective function of the traditional CNN.The regularization term emphasized the feature representation while also minimized the domain differences to achieve Cross-domain transfer of knowledge.Experimental results on public databases show that our method improved performance by 10%over traditional deep learning programs on cross-database VTR task.We propose a surveillance-based VTR method based on deep learning and active learning.By introducing an active learning algorithm framework to effectively annotate the vehicle images from the surveillance data,the pressure on large-scale data is effectively reduced.In real-world applications,the vehicle images set from surveillance data is lack of labels and redundant.Therefore,we select the high information entropy samples which were most beneficial to improve the performance of the recognition model,and reliable high confidence samples through the two-way threshold strategy into train set together.The high confidence samples provide new features for model learning and also soft the abrupt change of the original training data distribution to ensure the stable update of the model.Experimental results on a public database show that our method saves 40%of annotation cost when achieving the same level of reliable recognition ability compared with a randomly selection.