Search

Siamese-architecture-based self-supervised learning usually suffers from relatively high computational consumption and unsatisfactory performance because of its slow convergence and poor feature extraction capability. To alleviate these issues, we propose a self-supervised method, called SSL 2 , that is both efficient and effective. Specifically, a global and local sampling method is proposed to increase the number of samples while maintaining connections between semantic features. More significantly, SSL 2 maintains low computational complexity and facilitates the establishment of mapping relationships between global comprehensive and local detailed features compared with other methods. In addition, an i nformation r etainer p rojection h ead (IRPH) is proposed to further balance the information between detailed inconsistency and semantic consistency. Finally, hybrid tasks are embedded into SSL 2 to optimize the model so that it can effectively leverage the data provided by global and local sampling. Extensive qualitative and quantitative evaluations of various types of benchmarks illustrate that SSL 2 outperforms existing self-supervised frameworks in commonly used computer vision tasks. Specifically, SSL 2 achieved satisfactory performance with linear classification on ImageNet, outperforming MoCo-v2 by 2.2% with fewer calculations, and it also achieved competitive results compared with other state-of-the-art methods. • A new single-branch self-supervised learning framework is designed. • A novel sampling strategy is designed to make full use of the information. • A head is designed to balance detailed inconsistency and semantic consistency. • Well-balanced hybrid loss functions are introduced. [ABSTRACT FROM AUTHOR]

Convolutional neural network (CNN)-based self-supervised visual representation learning (SSL) is a long-standing problem achieving significant successes with traditional handcrafted pretext tasks and contrastive learning. However, existing SSL methods typically suffer from high computational overhead and poor performance due to sluggish convergence speeds and poor detail extraction capabilities. In this work, to address these issues and improve the robustness, we provide a new self-supervised architecture for incorporating a single-branch backbone with hybrid tasks into the representation learning process. Specifically, our method takes advantage of features from both intra- and inter-images by using discrete montage images. Then a single backbone with a novel A daptive D ecouple C onfusion (ADC) module is proposed to improve the feature extraction capabilities and alleviate the confusion regions in montage images. Besides, both concatenated discrete vectors and patch-based global average pooled vectors in latent space are utilized to learn local detailed features and maintain semantic consistency simultaneously. Moreover, our method is optimized by hybrid tasks and enjoys faster convergence speed due to these improvements. Extensive experiments on several datasets demonstrate the effectiveness and robustness of our method. The proposed method has 2.0% improved in linear classification to the conventional single-branch methods. [Display omitted] • A new framework SSH that leverages concatenated discrete and PGAP vectors is proposed. • An A daptive D ecouple C onfusion (ADC) module is proposed to avoid degenerate performance in existing jigsaw-based methods. • We demonstrate the advantages of the above contributions through extensive experiments. [ABSTRACT FROM AUTHOR]

Self-supervised visual representation learning focuses on capturing comprehensive features via exploiting the unlabeled datasets. However, existing contrastive learning based SSL frameworks are subjected to higher computational consumption and unsatisfactory performance. To handle these issues, we present a novel single-branch SSL method that incorporates an adaptive feature selection and activation module and a progressive cascade dimensionality reduction module, called APNet. Specifically, our method first fully exploits the unlabeled datasets and extracts intra- and inter-image information via introducing montage image. In addition, a novel adaptive feature selection and activation module is designed to generate the most comprehensive features. Besides, a progressive cascade dimensionality reduction module is proposed to capture the most representative features from latent vectors through cascade dimensionality increasing–decreasing operations. Extensive experiments have demonstrated the robustness and effectiveness of APNet. Specifically, APNet exceeds MoCo-v3 by 3.1% on the ImageNet-100 dataset, and consumes only half of the calculation. Code is available at https://github.com/AI-TYQ/APNet. • We embed feature selection and dimensionality reduction modules into SSL framework. • We design an attention module that selects and activates representative features. • We introduce a dimensionality reduction module to retain discriminative features. • We demonstrate the advantages of the above contributions through experiments. [Display omitted] [ABSTRACT FROM AUTHOR]