<p>Content-based image retrieval (CBIR) using visual saliency in the pixel domain has shown promising retrieval results at lesser computational cost as features are extracted only from salient regions. CBIR in the JPEG compressed domain has gained importance because these methods extract features without completely decompressing the image. However, no work on CBIR incorporates visual saliency models in the JPEG-compressed domain. In this paper, we propose a new method highlighting salient discrete cosine transform (DCT) blocks in the JPEG compressed domain to extract features. A saliency map is generated by applying a graph-based visual saliency model to the DCT coefficients of the image. Each DCT block of the image is categorized into salient and non-salient blocks according to its perceptual relevance, which is calculated by thresholding the saliency map. The feature vectors of different lengths from luminance and chrominance color components are extracted from salient DCT blocks. The impact of thresholding the saliency map, feature vector length, codebook size, and quantization parameter are also explored. The experimental results show that our proposed method incorporating visual saliency in the JPEG compressed domain gives comparable retrieval performance at a lesser computational cost when compared with state-of-the-art methods.</p>