Methods For Improving Generalization Of Convolutional Neural Networks On Image Classification And Detection

Convolutional neural networks have made great achievements in the field of computer vision.As the number of layers in the development of the network structure increases,the generalization performance of convolutional neural networks for unfamiliar data is increasingly difficult to understand.In this paper,starting from image classification and detection tasks,the current popularized convolutional neural network generalization theory is studied.The regularization methods of conventional neural networks are explored.Based on the original regularization methods,we carry out some methods.In order to expand and innovate,we explored the method of constructing a method similar to data augmentation directly on the convolutional layer of the neural network.Our experiments show that directly moving existing blending methods from classification to object detection will cause the training process become harder,eventually will lead to a bad performance.Inspired by our discovery,we presents a multi-phase blending method with incremental blending intensity to improve the accuracy of object detectors and achieve remarkable improvements.Firstly,to adapt blending method to detection task,we propose a smoothly scheduled and incremental blending intensity to control the degree of multi-phase blending.Based on the above dynamic coefficient,we propose an incremental blending method,in which the blending intensity is smoothly increased from zero to full.Therefore more complex and various data can be created to achieve the goal of regularizing the network.Secondly,we also design an incremental hybrid loss function to replace the original loss function.The blending intensity in our loss function increases smoothly,which is controlled by our scheduled coefficient.Thirdly,we further discard more negative examples in our multi-phase training process than other typical training methods and processes.By doing so,we can regularize the neural network to enhance generalization capability with data diversity and eventually to improve the accuracy in object detection.We also presents a novel method to exchange regions between a pair of feature maps.Therefore the diversity of feature maps can be greatly enhanced in addition to completely remove spatial correlated information from source feature maps for avoiding over-fitting.Then,corresponding to above exchange operation,we design a new loss function after extending the concept of the receptive field,which is transformed into a method to compute coefficients for the first time.Finally,in order to optimize the hybrid loss function,we define a fractional representation of receptive fields to account for the different weights of the edge and center pixels.The experiments on open benchmarks confirm that the proposed method outperforms other regularization methods for classification task.