BookCover Dataset Overview

• BookCover Dataset is a large-scale collection of Amazon book cover images with annotations for genre classification and document analysis.
• The benchmark includes a balanced BookCover30 subset with 57,000 images across 30 genres, facilitating deep learning and pattern recognition research.
• Extensions incorporating OCR-derived text and layout annotations enable multimodal fusion and synthesis studies, while exposing label ambiguities and intra-genre variance.

Searching arXiv for the primary and related BookCover dataset papers to ground the article in the cited literature. The BookCover Dataset denotes a family of datasets centered on book-cover imagery, with the canonical benchmark originating in “Judging a Book By its Cover” (Iwana et al., 2016). In its original form, it is a large-scale collection of Amazon.com book covers annotated by top-level genre and designed for visual genre classification. Subsequent work reused a balanced 30-class subset often called BookCover30, extended it with OCR-derived cover text for multimodal learning, and inspired distinct datasets that pair covers with summaries or add layout annotations for generative modeling (Kundu et al., 2020). As a result, the term refers both to a specific benchmark for genre prediction and, more broadly, to a line of book-cover datasets used in classification, multimodal retrieval, and text-to-image synthesis.

1. Origin, definition, and scope

The original BookCover Dataset was introduced as a challenging pattern-recognition benchmark for learning relationships between books and their covers (Iwana et al., 2016). The full dataset contains 137,788 unique book cover images in 32 classes, where each entry includes the cover image, title text, author text, and category membership. The categories follow Amazon’s top book categories, and each book is assigned to a single top-level category.

For deep learning experiments, the dataset was refined into a balanced subset of 30 classes, each containing 1,900 books, for a total of 57,000 images. Two genres, “Gay & Lesbian” and “Education & Teaching,” were excluded because they did not meet the 1,900-sample threshold, having 1,341 and 1,664 entries respectively (Iwana et al., 2016).

The 30 experimental genres are: Arts & Photography, Biographies & Memoirs, Business & Money, Calendars, Children’s Books, Comics & Graphic Novels, Computers & Technology, Cookbooks, Food & Wine, Crafts, Hobbies & Home, Christian Books & Bibles, Engineering & Transportation, Health, Fitness & Dieting, History, Humor & Entertainment, Law, Literature & Fiction, Medical Books, Mystery, Thriller & Suspense, Parenting & Relationships, Politics & Social Sciences, Reference, Religion & Spirituality, Romance, Science & Math, Science Fiction & Fantasy, Self-Help, Sports & Outdoors, Teen & Young Adult, Test Preparation, and Travel.

The dataset’s intended uses extend beyond genre prediction. The original paper explicitly positioned it for pattern recognition and deep learning experiments, document analysis, text recognition, font analysis, visual design analysis, feature extraction and representation learning, and the study of how features learned on natural images transfer to design and art domains (Iwana et al., 2016).

2. Construction, annotation, and preprocessing

The original corpus was collected from Amazon.com’s books section, with cover images and genre metadata obtained from listings (Iwana et al., 2016). Annotation was inherited directly from Amazon’s top-level classification. If a book belonged to multiple categories, one category was randomly chosen as its label, and no manual corrections or relabeling were made. For the balanced experimental subset, 1,900 books from each eligible class were randomly sampled.

This annotation scheme is central to the dataset’s character. Later analysis notes that many of the books belong to multiple genres, but only one genre was retained during preprocessing, and that this randomization “raises concerns on the validity of the dataset” and increases task complexity (Kundu et al., 2020). A plausible implication is that the benchmark combines genuine visual ambiguity with label noise induced by the single-label simplification.

Preprocessing in the original experiments was minimal. For AlexNet, images were resized to 227 × 227 × 3 (RGB); for LeNet, they were resized to 56 × 56 × 3 (RGB). Apart from resizing, the dataset was used as is, and “No pruning of cover images and no class membership corrections were done” (Iwana et al., 2016).

The original experimental protocol used a 90% train / 10% test random split. On the 57,000-image subset, this corresponds to 51,300 training images and 5,700 test images, with 1,710 and 190 samples per genre respectively (Iwana et al., 2016). The dataset was released publicly for academic use at https://github.com/uchidalab/book-dataset, while the cover images remained copyright Amazon.com, Inc., used under fair use for academic purposes.

3. BookCover30 as the standard benchmark

The balanced 30-class subset became the de facto benchmark for subsequent work and is explicitly referred to as the BookCover30 dataset by Iwana et al. (2016) in later multimodal studies (Kundu et al., 2020). In that setting, the dataset contains 57,000 book cover images, 30 genres, and exactly 1,900 samples per genre, preserving class balance.

A later benchmark protocol split the data into 80% training, 10% validation, and 10% test (Kundu et al., 2020). This yields 45,600 training images, 5,700 validation images, and 5,700 test images, with per-genre counts of 1,520, 190, and 190 respectively. All images were resized to 224 × 224 pixels so that they had the same dimension.

This re-standardization matters because it turns BookCover30 into a more typical modern benchmark with an explicit validation partition, enabling architecture selection and multimodal ablation. At the same time, the underlying semantic problem remains unchanged: book covers are graphic designs whose colors, styles, and textual information vary widely even within the same genre, while genre boundaries themselves are not always concretely defined (Kundu et al., 2020).

The benchmark’s difficulty is also tied to the fact that cover design varies with country, culture, and target reader populations, and can be pushed toward sales-oriented visual novelty. This suggests that the dataset captures not only genre semantics but also market-facing design conventions and their exceptions.

4. Multimodal extension through OCR and cover text

A major extension of the benchmark was introduced in “Deep multi-modal networks for book genre classification based on its cover” (Kundu et al., 2020). That work added an explicit text modality by extracting all visible texts on covers with the Google Cloud Vision API. The extracted text typically includes book titles, author names, publisher info, and sometimes additional text.

The extracted texts were then encoded using standard tokenization: each word was tokenized and assigned a unique integer ID, but only words appearing at least 5 times in the whole dataset were encoded. This threshold was used to reduce noise and dimensionality (Kundu et al., 2020). The authors emphasized that this differs from earlier work that used only title information stored in the original dataset; their approach used all the texts automatically extracted from the cover image.

This multimodal variant clarifies an important distinction. The original BookCover Dataset is not an OCR-annotated multimodal corpus in its base release; rather, it provides cover imagery plus metadata such as title and author. The OCR-derived text channel is a later augmentation built on top of the same visual benchmark (Iwana et al., 2016, Kundu et al., 2020).

The multimodal study also identified structural issues affecting fusion methods. It reported low inter-genre variance and high intra-genre variance, as well as cases where images and texts on the same cover suggest different genres. These conditions are particularly relevant for methods such as DCCA, which assume meaningful correlation between modalities (Kundu et al., 2020). This suggests that the BookCover30 benchmark is not merely multimodal; it is multimodal under partial inconsistency.

5. Reported performance and benchmark difficulty

On the original visual-only benchmark, AlexNet achieved 24.7% Top-1, 33.1% Top-2, and 40.3% Top-3 accuracy, compared with 13.5% Top-1 for LeNet (Iwana et al., 2016). The original paper highlighted the task’s difficulty: genre is not always visually evident, covers can be ambiguous or misleading, and the ground truth is intrinsically multi-label even though the benchmark is framed as single-label classification.

Per-genre performance was highly uneven. In the reported examples, Calendars reached 47.9% Top-1 and 65.3% Top-3, Test Preparation reached 68.9% Top-1 and 78.4% Top-3, Children’s Books reached 42.1% Top-1 and 61.6% Top-3, and Romance reached 45.3% Top-1 and 60.5% Top-3, while the overall average remained substantially lower (Iwana et al., 2016). A plausible implication is that some genres expose strong recurring design priors, whereas others are visually diffuse or overlap heavily with neighboring genres.

In the later multimodal benchmark, the best image-only result reported was ResNet-50 with 29.6% Top-1 and 49.0% Top-3, while the best text-only result, based on USE, reached 52.6% Top-1 and 73.7% Top-3 (Kundu et al., 2020). The best overall result came from a simple concatenation multimodal model, which achieved 56.1% Top-1 and 77.7% Top-3, outperforming the DCCA multimodal model at 48.9% Top-1 and 72.3% Top-3.

These numbers are frequently interpreted as evidence that visible cover text is highly informative for genre prediction. The reported analysis supports that interpretation, but it also emphasizes that more efforts and resources are needed to reach a satisfactory level of performance (Kundu et al., 2020). The benchmark therefore functions less as a solved classification task than as a stress test for weakly aligned visual and textual semantics.

6. Later variants, related datasets, and naming ambiguity

In later literature, “BookCover Dataset” does not always denote the original Amazon genre-classification corpus. At least two later resources reuse the name in different senses: a Goodreads-based summary-cover dataset for text-to-image synthesis, and a layout-annotated pedagogical BookCover dataset for STEM illustration generation (Haque et al., 2022, Zhang et al., 15 Aug 2025). Distinguishing these resources is essential because they differ in source, supervision, and intended task.

Resource	Source	Core properties
Original BookCover / BookCover30	Amazon.com	137,788 covers in 32 classes; 57,000-image balanced 30-class subset; cover, title, author, category (Iwana et al., 2016)
Goodreads summary-cover dataset	Goodreads	24,786 English books; summaries paired with cover images; 6 balanced genres (Haque et al., 2022)
Pedagogical BookCover for LEARN	Kaggle book covers + STEM textbooks	30,000+ images; bounding boxes, object labels, narrative relations for layout-to-image generation (Zhang et al., 15 Aug 2025)

The Goodreads dataset introduced for “Book Cover Synthesis from the Summary” contains 24,786 unique entries, retains only English-language books, and organizes data so that images and corresponding summaries have matching filenames (Haque et al., 2022). It is genre-balanced across Children, Mystery and Thriller, Non-fiction, Romance, Science Fiction, and Young Adult, and was explicitly designed for text-to-image synthesis rather than cover-based genre recognition. Its preprocessing included retaining summaries of at least 40 characters, creating multiple captions per image, and resizing cover images to dimensions such as 128×128 or 256×256 pixels for model training.

The LEARN framework describes another dataset called BookCover, curated from the public Book Covers dataset on Kaggle and enriched with additional illustrations from STEM textbooks, yielding a corpus exceeding 30,000 images (Zhang et al., 15 Aug 2025). In that resource, each image is annotated with bounding boxes, object labels $o_i$, and, for some images, object relationship annotations or semantic graphs, with the annotation format

$L = \{(o_i, b_i)\}_{i=1}^N,$

where $b_i \in \mathbb{R}^4$ gives normalized top-left coordinates and width/height. Its purpose is layout-to-image generation in educational contexts, not genre classification.

A common misconception is therefore to treat all “BookCover” datasets as interchangeable. They are not. The original benchmark is a genre-labeled visual corpus from Amazon (Iwana et al., 2016); the later Goodreads resource is a summary-to-cover paired dataset for synthesis (Haque et al., 2022); and the LEARN resource is a layout-annotated pedagogical corpus (Zhang et al., 15 Aug 2025). Another misconception is to assume that “BookCover” inherently means text-summary pairing. In fact, the canonical dataset was originally constructed for image-based genre prediction, and cover-text multimodality was added later through OCR (Kundu et al., 2020).

7. Research significance and open issues

The original BookCover benchmark established that computers can learn nontrivial relations between covers and genres from visual clues alone, but it also exposed the limits of single-image classification for a semantically messy cultural artifact (Iwana et al., 2016). Later multimodal work showed that extracting text from the cover itself materially improves performance, indicating that genre information is embedded not only in color palettes, illustration styles, and typography, but also in visible lexical content (Kundu et al., 2020).

At the same time, the benchmark encodes several unresolved issues. The randomized single-genre assignment for books that belong to multiple genres introduces label ambiguity; low inter-genre variance and high intra-genre variance make purely visual discrimination unstable; and inconsistent visual and textual information weakens methods that rely on strong cross-modal correlation (Kundu et al., 2020). These are not incidental nuisances but defining properties of the task.

For that reason, BookCover remains valuable less as a conventional clean-label benchmark than as a controlled testbed for weak supervision, multimodal fusion, OCR-enhanced document understanding, and the analysis of design rules, object dependencies, and color/font associations in cultural graphics (Iwana et al., 2016). The proliferation of later datasets that reuse the name further underscores its influence: the original benchmark became a reference point from which classification, synthesis, and layout-aware generation tasks all diverged in different directions.