Research On Object Recognition In Infrared Images

Recently,deep learning based architectures,notably Convolutional Neural Networks(CNNs),have enabled rapid advances in the field of artificial intelligence in general,and in the area of visual recognition in particular.This progress has triggered many real-world applications such as face detection and recognition,personal photo search,perception in robotics,and self-driving cars.However,all these computer vision based systems collapse in the case of poor lighting conditions or at the night time.Therefore,in this dissertation,we aim to develop algorithms for object recognition systems able to perform properly in the day and the night time using infrared thermal images.The proposed approach uses machine learning algorithms,especially convolutional neural networks and extreme learning machines,to learn distinctive features and perform accurate and fast classification.First,motivated by the rich hierarchical representations learned by the convolutional structures,we developed a model based on convolutional neural networks to learn useful features suitable for infrared thermal images without any explicit hand-engineering.Second,to enhance the accuracy of a single model,we presented a fully learnable ensemble of Extreme Learning Machines(ELMs)for multi-class classification,named ELM-In-ELM.The proposed approach learns automatically how to combine different individual models,using the ELM algorithm,by minimizing not only the classification error but also the norm of the network weights,which leads to better generalization performance.Moreover,interestingly,many ELM based ensembles can be regarded as special cases of the proposed ELM-In-ELM.Extensive experiments on 32 standard classification benchmarks including small and largesize/high-dimension datasets have been carried out.In addition,a comparison with different state-of-the-art models has been performed,which shows that the proposed unified ELMIn-ELM ensemble can achieve competitive results in term of generalization performance,network complexity,and training speed.Third,to handle the lack of training data and the problem of overfitting,a new fast and accurate approach based on extreme learning machine to train any convolutional neural network,named ELM-CNN,is proposed.The developed framework is based on the concept of auto-encoding to learn the convolutional filters with biases,by reconstructing the normalized input and the intercept term.Moreover,systematic comparison with traditional back-propagation based training method has been made with respect to two aspects qualitative and quantitative.The experimental results show that the proposed ELM-CNN algorithm achieves competitive results in term of generalization performance and run much faster than the back-propagation based training of convolutional neural networks.However,when the amount of available training data is small,the traditional backpropagation algorithm fails due to the problem of overfitting,while the proposed ELM-CNN algorithm achieves the best results in term of both the generalization performance and the training speed.Lastly,to show the effectiveness of the proposed algorithms in real-world problems,we have proposed two applications;pedestrian detection and maritime ships recognition.Indeed,we proposed a model based on convolutional neural networks to learn useful features to perform pedestrian detection using infrared thermal images.The developed pedestrian detector plays an important role in the advanced driver assistance systems that can perform adequately in both the day and the night time.In the case of maritime ships recognition,we combined the ELM based learning algorithm to train convolutional neural networks(ELMCNN)for discriminative features extraction and the ELM based ensemble for classification(ELM-In-ELM).The experimental results confirm that the proposed approach outperforms the state-of-the-art models in term of generalization performance and training speed.The proposed algorithms can extract distinctive representations using the ELM-CNN algorithm,and also perform fast and accurate classification using the ELM-In-ELM algorithm,which mean that the proposed approach can be applied in any object recognition system.