Deep neural network are one of the most powerful model for machine learning, which can learn underlying patterns automatically from a large amount data. So it be extensively used in more and Internet-of-Things (IoT) applications. However, training deep models is difficult, suffering overfitting gradient vanishing problem. Besides, parameters multiplication operations make impractical learning to directly execute on target hardware. In this paper, we propose method gradually pruning weakly...

The Domain Generalizable Re-identification (DG ReID) task has attracted significant attention in recent years, as a challenging but closely aligned with practical applications. Mixture-of-experts (MoE) based methods been studied for DG ReID to exploit the discrepancies and inherent correlations between diverse domains. However, most of methods, especially MoE-based have full fine-tune large amount parameters, which is not always real-world scenarios. Considering this problem, we propose...

One kind of Deep Learning models-convolutional neural network, which can reduce the complexity network structure and number parameters to be determined through local receptive fields, weight sharing pooling operation has achieved state art results in image classification problems. But this model gradient diffusion problem, cause slow updating underlying during process training. To solve problem above make improvements, paper presents a convolutional based on principal component analysis...

Cross-age face recognition problem is of great challenge in practical applications because features the same person at different ages contain variant aging addition to invariant identity features. To better extract age-invariant hiding beneath age-variant features, a deep learning-based approach with multiple attention mechanisms proposed this paper. First, we propose stepped local pooling strategy improve SE module. Then by incorporating residual-attention mechanism, self-attention and...

Object detection in the 2D domain is well developed owing to wide application of CMOS image sensors and great success deep learning technologies recent years. However, under circumstances such as autonomous driving, variation weather conditions light makes it impossible perform reliable using regular sensors. 3D data generated by a Lidar or Radar more robust environments, hence serving an essential complement scenarios. Well-established anchor-based detectors suffer from time-consuming...

For a learning automaton, proper configuration of the parameters is crucial. To ensure stable and reliable performance in stochastic environments, manual parameter tuning necessary for existing LA schemes, but procedure time-consuming interaction-costing. It fatal limitation LA-based applications, especially those environments where interactions are expensive. In this paper, we propose parameter-free automaton (PFLA) scheme to avoid by Bayesian inference method. contrast schemes must be...

Predicting bounding box with higher intersection over union (IoU) is one of the most important issues in many computer vision tasks. The ℓn-norm loss and IoU-based are two conventional approaches to guide a training process recent methods. However, optimization direction not exactly same as maximizing metric. In addition, suffers from some inevitable disadvantages due direct addition IoU. According shape, size, position properties, we design mixed geometric (MG) regression increase...

The combination of a CNN detector and search framework forms the basis for local object/pattern detection. To handle waste regional information defective compromise between efficiency accuracy, this paper proposes probabilistic model with powerful framework. By mapping an image into distribution objects, new gives more informative outputs less computation. setting analytic traits are elaborated in paper, followed by series experiments carried out on FDDB, which show that proposed is sound,...