Transfer Learning Based Multi-class Object Recognition And Detection

Category-level object recognition and detection are fundamental problems in computer vision,which aims to study the problem of identifying and localizing interested objects in a given image or video stream.For category-level object recognition and detection tasks,the key issues and challenges are interwoven,such as overfitting,class distribution imbalance and cross-domain model adaptation.When we have inadequate training data,the trained model tends to overfit the training data,which means that it performs well on the training data but its generalization ability is very poor on the test data.In addition,when we have imbalanced training data for different object categories,the trained model tends to bias toward the majority class.Since the feature distributions extracted from different types of images are very different,the performance of a model trained on a source domain may be significantly degraded on the target domain for cross-domain object recognition and detection tasks.In this paper,we apply the theory of transfer learning to solve the problems encountered in category-level object recognition and detection application,and experimentally demonstrate the effectiveness of the proposed algorithms when performing multiclass object recognition and detection tasks.The main works are given as follows:For the binary class object recognition problem under the condition of small training data and class imbalance,we propose a novel method named WSMOTE to generate cross domain weighted synthetic instances,which assigning a proper weight for each generated synthetic instance and dynamically incorporate them into the training process of a weak classifier at each iteration of the TrAdaboost algorithm.Based on WSMOTE method and TrAdaboost learning framework,we also propose a novel WSMOTE-TrAdaboot algorithm,which can selectively transfer a part of beneficial instances from the source and synthetic instances to the target domain.After the instance-level transfer learning,the number of positive and negative training data in the target domain can be rebalanced.For the feature distribution shift problem in large-scale cross-domain multiclass object recognition application,we proposed three online feature transformation learning algorithms and prove their theoretical loss bounds.The learning of feature transformation is viewed as online learning a cross-domain similarity metric function,and then use the learning similarity function in the k Nearest Neighbor(k-NN)classifier to perform multiclass object recognition task.In addition,a novel method is also proposed to incrementally fine-tune Convolve Neural Networks(CNNs)and then combine these fine-tuned CNN models for the same application.When we only have some pre-trained source CNN models and few target training data,the proposed method can be used to alleviate the CNN model overfiting risk for cross-domain and multiclass object recognition task.For the problems encountered in the multiclass object category detection by a mobile robot in cluttered indoor scene,we design and implement a 3D object detection system by sharing fragment features of bearing angle image.We first build a novel 3D point clouds data set in cluttered indoor scene named DUT-3D by a 3D laser scanner,and then we jointly train multiple object detectors by sharing fragment features of bearing angle image manner.Next,the reclassification technique and HOG-LBP features are used to reclassify the.detected low confidence bounding boxes.A novel RUS-SMOTEboost algorithm is proposed for the reclassification task,which can be used to discriminatively train a group of binary classifiers with class imbalanced training data.