<p>Contrastive learning (CL) has been widely adopted in deep quantization-based image retrieval. Existing approaches typically generate two views of an image via traditional data augmentation and construct negative pairs using different images. However, randomly selected negatives often introduce false-negative samples, which hinder model training and degrade retrieval performance. Additionally, relying solely on global features fails to capture fine-grained similarity in deep quantization-based retrieval. To address these issues, a novel framework, Negative view-based Contrastive Deep Accumulative Quantization (NCDAQ) for unsupervised image retrieval, is proposed. NCDAQ introduces a patch-based negative augmentation strategy that disrupts the semantic integrity in input images to generate meaningful true-negative views, forming accurate negative pairs for CL. Moreover, an accumulative quantization (AQ) is integrated with negative view-based contrastive learning to approximate the image representation vector more accurately. To further enhance the discrimination of image representation, a global–local similarity consistency module is employed to reduce background noise and emphasize the feature in target regions. Extensive experiments on three public datasets demonstrate that NCDAQ significantly outperforms state-of-the-art methods in image retrieval.</p>