Source-linked AI summary
Deep Learning for Logo Recognition
Simone Bianco, Marco Buzzelli, Davide Mazzini, Raimondo Schettini
TL;DR
Logo recognition must handle logos appearing at varied positions, scales, and viewpoints, as well as image artifacts and distortions. This paper combines candidate-region proposals with a CNN trained for logo classification, then reports that its best solution outperforms prior methods.
Problem
Logos can appear at any position, scale, or point of view and images may contain artifacts and distortions.
Source: §5 · p. 8Method
The pipeline feeds recall-oriented logo-region proposals to a CNN specifically trained for classification, including when regions are imperfectly localized.
Source: §5 · p. 8Results
The best solution outperforms state-of-the-art methods, improving F1-measure by 3.8% and accuracy by 1.7% over the respective best prior methods.
Source: §4 · p. 8, §5 · p. 10Takeaways & Limitations
The best-performing configuration uses a background class, precise and object-proposal annotations, synthetic augmentation, epoch-based class balancing, and contrast normalization, with sample weighting disabled.
Source: §5 · p. 10
Abstract
from arXiv · showhide
In this paper we propose a method for logo recognition using deep learning. Our recognition pipeline is composed of a logo region proposal followed by a Convolutional Neural Network (CNN) specifically trained for logo classification, even if they are not precisely localized. Experiments are carried out on the FlickrLogos-32 database, and we evaluate the effect on recognition performance of synthetic versus real data augmentation, and image pre-processing. Moreover, we systematically investigate the benefits of different training choices such as class-balancing, sample-weighting and explicit modeling the background class (i.e. no-logo regions). Experimental results confirm the feasibility of the proposed method, that outperforms the methods in the state of the art.
1. Introduction
Earlier logo recognition relied on keypoint-based methods, while more recent work explored deep learning with pretrained CNNs and synthetic data. This paper proposes a deep-learning logo-recognition method and investigates training choices for the task.
1. Introduction
- Earlier logo-recognition methods used keypoint-based detectors and descriptors, including feature bundling, spatial-layout encoding, burstiness modeling, and homographic class graphs.Some methods also used inverted secondary models to handle logos with inverted colors.Source: §1 · p. 1
- Recent deep-learning approaches explored pretrained CNNs, synthetic data, and different network architectures for logo recognition.A related pipeline used pretrained CNNs with Fast Region-Based Convolutional Networks.Source: §1 · p. 1
- The proposed method uses deep learning for logo recognition and investigates image preprocessing, class balancing, sample weighting, synthetic augmentation, and a background class.It also adds object-proposal regions as positive training examples and examines real-data augmentation.Source: §1 · p. 1
2. Proposed Method
The method proposes candidate logo regions and classifies them with a CNN trained to handle imperfect localization. Its training framework tests annotation sources and training choices, while testing aggregates proposal predictions and applies a learned threshold.
2. Proposed Method
- Training can use precise ground-truth logo crops alone or combine them with overlapping object-proposal regions labeled with class and IoU.The framework investigates class balancing, augmentation, contrast normalization, sample weighting, and a background class formed from proposals that do not overlap logos.Source: §2 · p. 2
- At test time, proposal predictions are max-pooled, and the highest-confidence class is assigned only if it exceeds a learned threshold.The same proposal algorithm is used for training and testing; contrast normalization is applied at test time if enabled during training.Source: §2 · p. 2
2.1. Object proposal
- Selective Search supplies recall-oriented candidate regions, accepting many false positives so the CNN can disambiguate them.Proposals are cropped to the CNN input size; training proposals overlapping a ground-truth logo by IoU≥0.5 can be used as positive examples.Source: §2 · pp. 1–2, §2.1 · p. 2
2.2. Network Architecture
- The classifier is a tiny CNN with three convolutional layers and two fully connected layers, containing 1.5 × 10^5 parameters.Its small size is intended to support fast testing and training on inexpensive GPUs; the architecture also uses ReLU, pooling, and a final Softmax classifier.Source: §2.2 · p. 4
3. Logos datasets
The paper uses FlickrLogos-32 and introduces Logos-32plus, a larger collection intended to broaden the range of logo appearances and imaging conditions. It describes dataset collection and a semi-automatic procedure for removing duplicates.
3.1. FlickrLogos-32 Dataset
- FlickrLogos-32 contains photos from 32 logo brands, with 10 training, 30 validation, and 30 test images per class.It is intended for evaluating logo retrieval and multi-class logo detection and recognition on real-world images.Source: §3.1 · p. 4
3.2. Logos-32plus Dataset
- Logos-32plus contains 12,312 instances and averages 400 examples per class; each image includes one or multiple instances of the same class, with images retrieved from Flickr and Google image search.The collection aimed to account for a larger set of real imaging conditions and transformations than Logos-32, whose images were usually focused, non-blurry or noisy, and highly saturated.Source: §3.2 · p. 4
3.3. Duplicates Removal
- Duplicate removal first discarded image pairs with SSIM above 0.9, then manually checked near duplicates among the classifier’s five nearest results.Near duplicates include images of the same scene with small appearance differences, such as overlapping crops or views from different points of view.Source: §3.3 · pp. 4–5
4. Experimental Setup and Results
The best configuration, TC-VII, improves on the strongest state-of-the-art results in both F1-measure and accuracy. Training choices produce large gains over the baseline, while GPU memory-transfer overhead makes the tested CPU solution faster overall.
4. Experimental Setup and Results
- Adding background examples improves F1-measure by 31.8% and accuracy by 52.4% over TC-I.Adding Selective Search proposals then improves F1-measure by 11.3% and accuracy by 12.4% over TC-II.Source: §4 · pp. 5, 7
- Synthetic proposal augmentation improves F1-measure by 11.7% and accuracy by 20.9% over TC-III.It increases the number of Selective Search proposals by perturbing them with random translations.Source: §4 · p. 7
- Epoch class balancing improves F1-measure by 0.4% and accuracy by 0.3% over TC-IV; contrast normalization adds 2.4% F1-measure and 0.2% accuracy over TC-V.Epoch balancing performs consistently better than batch balancing; overlap-based sample weighting lowers final performance.Source: §4 · p. 7
- The proposed solution improves F1-measure by 3.8% and accuracy by 1.7% over the best state-of-the-art methods.The strongest prior F1-measure and accuracy results came from different methods.Source: §4 · pp. 7–8
- Training on FlickrLogos-32 alone drops F1-measure by 14.7% and accuracy by 4.8%; adding ground-truth regions raises them by 0.6% and 0.2%, respectively.The authors interpret the small ground-truth-region gains as indicating that the CNN, rather than Selective Search, is the major error source.Source: §4 · p. 8
4.1. Timings
- The tested GPU solution is slower overall than CPU classification because CPU-to-GPU memory transfer adds overhead.Proposal extraction always runs on CPU; timings averaged 100 runs on different images.Source: §4.1 · p. 8
5. Conclusions
Logos vary in position, scale, and viewpoint and may suffer image distortions. The authors use recall-oriented region proposals with a task-trained CNN and report that their best configuration outperforms state-of-the-art methods.
5. Conclusions
- A recall-oriented region proposal feeds a task-trained CNN that classifies logos despite imperfect localization, addressing logos’ varying position, scale, viewpoint, and image distortions.Source: §5 · p. 8
- Figure 8 orders wrongly labeled logos by confidence at 32x32 resolution and pairs them with nearest training examples using pre-softmax activations as features.Source: Figure 8 · p. 9
- On FlickrLogos-32 and Logos-32plus, the best configuration outperforms methods in the state of the art.Table 4 compares the best configuration with state-of-the-art methods.Source: Table 4 · p. 9, §5 · p. 10
- The best configuration uses explicit background modeling, precise and actual object-proposal logo annotations during training, synthetic augmentation, epoch-based class balancing, and contrast normalization, with sample weighting disabled.Source: §5 · p. 10