Despite the evolution of deep-learning-based visual-textual processing\nsystems, precise multi-modal matching remains a challenging task. In this work,\nwe tackle the task of cross-modal retrieval through image-sentence matching\nbased on word-region alignments, using supervision only at the global\nimage-sentence level. Specifically, we present a novel approach called\nTransformer Encoder Reasoning and Alignment Network (TERAN). TERAN enforces a\nfine-grained match between the underlying components of images and sentences,\ni.e., image regions and words, respectively, in order to preserve the\ninformative richness of both modalities. TERAN obtains state-of-the-art results\non the image retrieval task on both MS-COCO and Flickr30k datasets. Moreover,\non MS-COCO, it also outperforms current approaches on the sentence retrieval\ntask.\n Focusing on scalable cross-modal information retrieval, TERAN is designed to\nkeep the visual and textual data pipelines well separated. Cross-attention\nlinks invalidate any chance to separately extract visual and textual features\nneeded for the online search and the offline indexing steps in large-scale\nretrieval systems. In this respect, TERAN merges the information from the two\ndomains only during the final alignment phase, immediately before the loss\ncomputation. We argue that the fine-grained alignments produced by TERAN pave\nthe way towards the research for effective and efficient methods for\nlarge-scale cross-modal information retrieval. We compare the effectiveness of\nour approach against relevant state-of-the-art methods. On the MS-COCO 1K test\nset, we obtain an improvement of 5.7% and 3.5% respectively on the image and\nthe sentence retrieval tasks on the Recall@1 metric. The code used for the\nexperiments is publicly available on GitHub at\nhttps://github.com/mesnico/TERAN.\n