Research On Feature Coding Method For Image Classification

At present,the degree of informatization in all walks of life is deepening,information resources show a blowout growth trend.How to effectively deal with massive data resources has become a focus of attention from all circles.As an important research direction in the field of computer vision,image classification plays a crucial role in image resource processing.In the framework of image classification,feature coding plays a role of connecting the past and the future,and has a vital impact on the accuracy and time of classification.This paper mainly focuses on the feature coding technology in image classification.The specific research work is as follows:(1)Considering that in face recognition,the sparse representation algorithm is time-consuming,and the collaborative representation algorithm is less robust against noise,occlusion and other interference.This paper proposes a singular vector coding method for face recognition.This method encodes the unknown sample as a linear combination of all known samples,which obtains singular vectors by singular value decomposition of each class of known samples,and uses singular vectors to encode features to obtain image representation.The experimental results on three common face datasets show that this method has high recognition accuracy and good robustness,and the computation time is far less than that of the sparse representation algorithm.(2)Aiming at the quantization error that may be caused by the hard coding strategy in the coding process,and the problem that the soft coding strategy is difficult to select the best number of visual words for feature descriptors,an improved Vector of Locally Aggregated Descriptors(VLAD)coding method for scene image classification is proposed.The new method extends the traditional soft coding strategy,calculates the distance variance of the K-nearest visual words of each descriptor to determine the number of visual words.In addition,it assigns weights to these visual words utilizing saliency.This paper has carried out relevant experiments on three benchmark datasets.The experimental results show that compared with the original VLAD method,the improved VLAD has better classification performance,and there is no significant change in computing time.(3)Aiming at the problems of low complexity of training tasks and insufficient negative samples in current contrastive self-supervised learning methods,this paper proposes a multi-network contrastive learning method for large-scale image classification.This method tries to use three networks for two-by-two contrastive learning to increase the task complexity.As many negative samples as possible are provided by combining the same batch of sample sets and the sample library maintained by the momentum encoder.This paper conducts relevant experiments on three common deep learning datasets.It is proved that the two proposed improvement measures are helpful to improve the performance of the model,and the classification performance of the new method is better than the existing baseline method.