Due to the high cost of manual annotation, unsupervised deep hashing has recently attracted much attention from researchers for large-scale image retrieval. However, the existing methods usually leverage Contrastive Learning to learn the image hashing codes, which suffers from one key issue of false negatives when training a model. Specifically, false negatives of model training may misclassify similar images as distinct ones, resulting in inconsistent hash codes. To address the challenges of unsupervised deep hashing, we propose a novel method termed Self-Quantization with Adaptive Codebooks (SQAC). SQAC directly leverages soft quantized features for self-supervised learning without negative samples. Furthermore, an adaptive codebook mechanism is proposed to enhance the codebook diversity and minimize the generalization error. The proposed SQAC is designed to derive a sparser codebook distribution suitable for constructing more discriminative image hashing codes. The extensive experimental results on four benchmark image datasets demonstrate the efficacy of the proposed SQAC method.