AI Reads the Mood of Your Restaurant Photos With Wavelet-Boosted Deep Learning

A plate of pasta can lie in a caption, but the photograph rarely does. That intuition, familiar to anyone who scrolls through Yelp or TripAdvisor before booking a table, is now the backbone of a new artificial intelligence framework that claims to read sentiment directly from images. In a study published in Discover Artificial Intelligence, researchers led by Siddhi Kadu and Bharti Joshi of the Ramrao Adik Institute of Technology, together with Pratik Agrawal of Symbiosis Institute of Technology in Nagpur, describe a pipeline that classifies restaurant review images as positive, negative or neutral, and outperforms a series of established deep learning baselines in the process.

The shift the researchers target is easy to overlook but fundamental to how the modern web works. Sentiment analysis began as a text problem: algorithms read reviews and scored words for polarity. But smartphones changed the medium of feedback. The paper notes that more than 70 percent of social media posts are visual, and that roughly 60 percent of readers prefer content enriched with images. Reviewers increasingly post photographs of food presentation, dining room ambiance and their own facial expressions, and those images carry emotional signals that text-only systems simply cannot access. A glowing written review paired with a photo of a wilted salad tells a more complicated story than either element alone.

Existing visual sentiment analysis approaches, the authors argue, share two weaknesses. First, most rely exclusively on convolutional neural networks that extract spatial features, meaning patterns in pixel arrangements such as objects, textures and edges. Spatial features work well for detecting a smiling face or a plated dish, but they miss subtler cues, like fine texture variations in food surfaces or gradual color gradients in ambient lighting, that live in the frequency domain of an image. Second, most pipelines skip a rigorous feature selection step, feeding classifiers an inflated vector that contains redundant or uninformative attributes. That redundancy inflates computational cost and can quietly degrade accuracy.

The proposed answer, named WaveNet-CNN, is a dual-domain feature extractor. One branch is a custom convolutional neural network with six convolutional layers, each followed by batch normalization and ReLU activation, with filter counts stepping up from 32 to 512 as the network climbs from low-level cues like edges toward high-level semantics like faces and emotions. Max-pooling layers shrink the feature maps between blocks, and a global average pooling operation condenses everything into a compact 512-dimensional vector. The architecture borrows the simplicity of VGG-style designs while avoiding bulky fully connected layers, a choice the authors say reduces overfitting while preserving expressive power.

The second branch is more unusual: an optimal wavelet transform tuned specifically for sentiment analysis. Wavelet transforms decompose an image into frequency subbands, separating the approximation of the image from horizontal, vertical and diagonal detail components. Instead of using fixed standard filters such as Daubechies or biorthogonal wavelets, the team constructed an Optimal Wavelet Filter Bank whose low-pass and high-pass analysis coefficients were modified at three settings of vanishing moments, v equals 2, 4 and 6. The coefficients were chosen by minimizing their Canonical Signed Digit representation, which reduces arithmetic complexity while retaining the properties that make wavelets useful, such as multi-resolution decomposition and frequency selectivity. In experiments, the highest setting, OWFB6, delivered the best trade-off between discriminability and efficiency.

Once the spatial vector from the CNN and the frequency vector from the wavelet bank are concatenated, the combined 1512-dimensional representation still contains plenty of dead weight. Enter the third ingredient: a Dual Moth Flame Optimization algorithm, an enhanced version of the nature-inspired Moth Flame Optimization metaheuristic. Conventional MFO evaluates candidate solutions, the moths, against a single fitness function. The dual variant instead scores each candidate using both intraclass variance, which measures how much a feature varies within each sentiment class, and interclass variance, which measures how sharply it separates different classes. The goal is to shrink variation within classes while widening the gap between them, yielding a compact subset of genuinely discriminative features. With 30 moths, 15 flames and 100 iterations, the algorithm trimmed 1512 features down to 725.

Classification is then handled by a weighted soft voting ensemble that combines five base learners: support vector machine, Naive Bayes, K-nearest neighbors, logistic regression and Deep Forest. Each base learner’s vote is weighted by its validation accuracy, so stronger models carry more influence. The authors compared this scheme against soft voting, hard voting, bagging, AdaBoost, gradient boosting and stacking, and report that the weighted ensemble consistently achieved the strongest results across the tested wavelet settings.

The evaluation spans three datasets. FER2013, a standard facial expression benchmark of roughly 35,887 grayscale images, was regrouped from seven emotion categories into positive, neutral and negative classes. JAFFE, a smaller controlled dataset of 213 images from 10 subjects provided by Kyushu University, was expanded to 1,704 samples for the experiments. Most importantly, the team assembled a restaurant review dataset of 935 positive, 1,007 neutral and 920 negative images curated from Yelp, TripAdvisor, Google reviews and prior research datasets. To improve generalization without leaking test data, the training split alone was augmented through horizontal flips, rotations up to 40 degrees, gamma contrast adjustments, Gaussian noise, sharpening, coarse dropout and blur, expanding it to more than 26,000 images. The WaveNet-CNN framework reached 75.15 percent accuracy on FER2013, 87.42 percent on JAFFE and 89.72 percent on the restaurant dataset, with an F1 score of 89.71 percent in the latter case, edging out DenseNet-121, EfficientNet-B0 and Swin Transformer-T baselines.

An ablation study traces where the gains come from, and the story is one of compounding effects. A standard CNN paired with logistic regression managed 71.50 percent accuracy on the restaurant data, while wavelet features alone with a biorthogonal db4-style bank reached only 72.65 percent. Swapping in the optimized OWFB6 filter bank lifted performance to 73.40 percent, and the custom CNN alone jumped to 78.84 percent. Combining custom CNN features with OWFB6 and the weighted ensemble pushed accuracy to 87.87 percent, and adding MFO-based selection raised it again to 88.76 percent before the full dual-objective DMFO pipeline landed at 89.72 percent. Statistical validation across ten independent runs, with paired t-tests yielding p-values below 0.001 and a 95 percent confidence interval of roughly 88.6 to 90.8 percent, supports the claim that the improvements are not artifacts of a lucky training run.

The framework is not without limits, and the authors are candid about them. Cross-dataset tests showed accuracy dropping to 66.30 percent when a model trained on facial expressions was applied to restaurant images and to 62.10 percent when trained on JAFFE and tested on the review data, revealing a substantial domain gap between controlled face datasets and messy real-world photography. The system also depends on image quality, handles only three sentiment classes rather than fine-grained emotions, ignores accompanying captions, and carries computational overhead from the wavelet decomposition and metaheuristic selection that makes it less suitable for real-time use. Future work, the team writes, will explore multimodal fusion with review text, attention-based and transformer architectures, comparison with vision-language foundation models such as CLIP, and explainability techniques like Grad-CAM and SHAP. Still, the practical payoff is concrete: restaurants could monitor customer sentiment in real time through review photos, spotting problems with specific dishes or decor before they snowball into reputation damage, and the approach points toward a web where machines understand not just what we write about our dinners, but what our cameras reveal about them.

Subject of Research: A deep learning framework combining CNN and optimal wavelet transform features for visual sentiment classification of restaurant review images

Article Title: Deep feature extraction based WaveNet CNN framework for visual sentiment analysis

Article References: Kadu, S., Joshi, B., & Agrawal, P. (2026). Deep feature extraction based WaveNet CNN framework for visual sentiment analysis. Discover Artificial Intelligence, 6(1), Article 1266. https://doi.org/10.1007/s44163-026-02185-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02185-0

Keywords: visual sentiment analysis, deep learning, convolutional neural network, wavelet transform, feature selection, moth flame optimization, ensemble classifier, restaurant reviews, computer vision, FER2013, JAFFE, image classification