Self‑Supervised Learning and Contrastive Learning for Computer Vision and OCR Applications

Self‑supervised learning addresses the lack of labeled data by using intrinsic data relationships as supervision, enabling the training of powerful feature encoders without explicit annotations. The article first outlines the four major machine‑learning paradigms—supervised, unsupervised, self‑supervised, and reinforcement learning—highlighting why self‑supervised methods have become a focus in recent years.

Typical computer‑vision pretext tasks are introduced, including predicting image rotation, solving jigsaw puzzles, patch location prediction, image colorization, auto‑encoding, and generative adversarial networks. For each task, the data preparation steps and the training objectives are described.

The core of modern self‑supervised methods is contrastive learning, which relies on constructing positive and negative sample pairs and designing appropriate loss functions. The article details several contrastive losses—Contrastive Loss, Triplet Loss, N‑Pair Loss, and InfoNCE Loss—providing their mathematical formulations and practical considerations such as avoiding model collapse.

Prominent contrastive frameworks are surveyed: SimCLR (large batch of negative samples), MoCo (momentum encoder with a queue of negative keys), and SimSiam (asymmetric architecture without negatives). Implementation snippets are provided, for example the momentum update of the key encoder:

@torch.no_grad()
for param_q, param_k in zip(self.encoder_q.parameters(), self.encoder_k.parameters()):
    param_k.data = param_k.data * self.m + param_q.data * (1. - self.m)

and the feature flattening used for sequence data:

def feature_flat(self, feature):
    dim = tf.shape(feature)[2]
    feature = tf.keras.layers.AvgPool1D(pool_size=5, padding="same")(feature)
    feature = tf.reshape(feature, [-1, dim])
    return feature

To illustrate practical impact, the article presents a case study on OCR for captcha recognition. Using 860k unlabeled captcha images for self‑supervised pre‑training (SimCLR and SimSiam) and a small labeled subset for fine‑tuning, the authors achieve a 5% accuracy improvement over purely supervised training (from 90.8% to 95.7%). Training details, hyper‑parameters, and a result table are included.

Finally, the article lists extensive references covering self‑supervised learning, contrastive methods, and related tools, and concludes with a brief thank‑you note.