Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Color-$S^{4}L$: Self-supervised Semi-supervised Learning with Image Colorization

Hanxiao Chen

TL;DR

This paper tackles semi-supervised image classification by integrating multiple self-supervised pretext tasks, notably image colorization, into a unified Color-$S^{4}L$ framework. It trains with two data streams—a labeled branch and an unlabeled, self-supervised branch—using a shared backbone and a multi-task loss $L_{Color-S^{4}L} = L_{super}(y_i,z_i) + \omega L_{self}( ilde{y_i},\tilde{z_i})$ to extract information from unlabeled data. Proxy labels arise from image rotation, geometric transformations, and a colorization task produced by an Encoder-Decoder model, including a colorization label $ ilde{y}=7$. Evaluations on CIFAR-10, SVHN, and CIFAR-100 across multiple CNN backbones show competitive or state-of-the-art results under limited-label conditions, underscoring the practical value of combining self-supervised colorization with SSL.

Abstract

This work addresses the problem of semi-supervised image classification tasks with the integration of several effective self-supervised pretext tasks. Different from widely-used consistency regularization within semi-supervised learning, we explored a novel self-supervised semi-supervised learning framework (Color-$S^{4}L$) especially with image colorization proxy task and deeply evaluate performances of various network architectures in such special pipeline. Also, we demonstrated its effectiveness and optimal performance on CIFAR-10, SVHN and CIFAR-100 datasets in comparison to previous supervised and semi-supervised optimal methods.

Color-$S^{4}L$: Self-supervised Semi-supervised Learning with Image Colorization

TL;DR

This paper tackles semi-supervised image classification by integrating multiple self-supervised pretext tasks, notably image colorization, into a unified Color- framework. It trains with two data streams—a labeled branch and an unlabeled, self-supervised branch—using a shared backbone and a multi-task loss to extract information from unlabeled data. Proxy labels arise from image rotation, geometric transformations, and a colorization task produced by an Encoder-Decoder model, including a colorization label . Evaluations on CIFAR-10, SVHN, and CIFAR-100 across multiple CNN backbones show competitive or state-of-the-art results under limited-label conditions, underscoring the practical value of combining self-supervised colorization with SSL.

Abstract

This work addresses the problem of semi-supervised image classification tasks with the integration of several effective self-supervised pretext tasks. Different from widely-used consistency regularization within semi-supervised learning, we explored a novel self-supervised semi-supervised learning framework (Color-) especially with image colorization proxy task and deeply evaluate performances of various network architectures in such special pipeline. Also, we demonstrated its effectiveness and optimal performance on CIFAR-10, SVHN and CIFAR-100 datasets in comparison to previous supervised and semi-supervised optimal methods.
Paper Structure (7 sections, 1 equation, 2 figures, 3 tables)

This paper contains 7 sections, 1 equation, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Overview of Color-$S^{4}L$ model
  4. Experiments Setup and Results
  5. Datasets and Model Architectures
  6. Color-$S^{4}L$ model Results and Analysis.
  7. Conclusions

Figures (2)

  • Figure 1: Color-$S^{4}L$ architecture for semi-supervised image classification. The left of pipeline denotes two kinds of self-supervised proxy tasks which include Image Colorization and Image Rotation. Also, we employ Geometric Transformation function $p(x)$ to produce 6 proxy labels defined as image rotation in multiples of 90 degrees ($[0^{\circ},90^{\circ},180^{\circ},270^{\circ}]$) along with horizontal(left-right) and vertical(up-down) flips like 23. In addition, we especially utilize Image Colorization function $h(x)$ to create the $7^{th}$ auxiliary label to strengthen the existing self-supervisions on unlabeled data within semi-supervised learning paradigm.
  • Figure 2: An overview of the Encoder-Decoder image colorization model architecture.