Source-linked AI summary
Towards Highly Accurate Coral Texture Images Classification Using Deep Convolutional Neural Networks and Data Augmentation
Anabel Gómez-Ríos, Siham Tabik, Julián Luengo, ASM Shihavuddin, Bartosz Krawczyk, Francisco Herrera
TL;DR
Coral texture patches omit whole-coral structure and can show similar species, variable appearances, and overlapping classes, making classification difficult. The paper evaluates CNN architectures with transfer learning and data augmentation, and reports state-of-the-art accuracy on EILAT and RSMAS, with ResNet best on both.
Problem
Coral texture images omit global structure and present similar species, within-class variation, and overlapping class boundaries, making expert-assisted classification challenging.
Source: §1 · p. 2Method
The study evaluates Inception v3, ResNet, and DenseNet on EILAT and RSMAS using ImageNet-initialized networks and examines data augmentation.
Source: §1 · p. 3, §5.1 · p. 13Results
CNN models achieved state-of-the-art accuracy on EILAT and RSMAS, with ResNet best on both and little benefit from data augmentation.
Source: §6 · p. 19Takeaways & Limitations
ResNet-based CNNs surpassed classical methods that require high human intervention on both coral datasets.
Source: §6 · p. 19Misclassified close-up images in both datasets include visually similar classes that may require expert distinction; whole-coral images may help for RSMAS.
Source: §5.4 · p. 19
Abstract
from arXiv · showhide
The recognition of coral species based on underwater texture images pose a significant difficulty for machine learning algorithms, due to the three following challenges embedded in the nature of this data: 1) datasets do not include information about the global structure of the coral; 2) several species of coral have very similar characteristics; and 3) defining the spatial borders between classes is difficult as many corals tend to appear together in groups. For this reason, the classification of coral species has always required an aid from a domain expert. The objective of this paper is to develop an accurate classification model for coral texture images. Current datasets contain a large number of imbalanced classes, while the images are subject to inter-class variation. We have analyzed 1) several Convolutional Neural Network (CNN) architectures, 2) data augmentation techniques and 3) transfer learning. We have achieved the state-of-the art accuracies using different variations of ResNet on the two current coral texture datasets, EILAT and RSMAS.
1. Introduction
Coral images are difficult to classify because species can look alike, images vary within classes, and texture patches omit whole-coral structure. The paper evaluates CNN architectures and data augmentation on coral datasets, comparing with classical methods.
1. Introduction
- Coral classification is difficult because species may appear identical, images vary within a species, and species can overlap in groups.Underwater illumination variation and changing taxonomy add to the difficulty of classification.Source: §1 · p. 2
- Earlier coral-image classification often combined machine-learning models with image enhancement or feature extraction, and few studies used multiple datasets.The paper identifies Shihavuddin et al. (2013) as the only cited work using several datasets.Source: §1 · p. 2
- The study evaluates Inception v3, ResNet, and DenseNet on EILAT and RSMAS texture patches and compares results with Shihavuddin et al. (2013).The patches show only part of a coral, not its entire structure.Source: §1 · p. 3
- The paper analyzes CNN performance, the impact of data augmentation, and comparisons with state-of-the-art classical methods.The authors frame the classical methods as requiring substantial human supervision and intervention.Source: §1 · p. 3
2. CNN Classification Models
This section introduces three CNN architectures and the optimization techniques used to address the small size of the coral datasets. It describes their building blocks and the paper’s transfer-learning and augmentation approaches.
2. CNN Classification Models
- The section situates Inception v3, ResNet, and DenseNet as the three architectures considered, alongside techniques intended to address small datasets.The architecture overview also describes increasing network depth, width, and reuse of lower-layer features.Source: §2 · pp. 3–4
2.1. Inception v3
- Inception v3 modifies GoogLeNet’s base module by replacing its 5×5 convolution with two 3×3 convolutions and factorizing deeper convolutions to reduce computational cost.The paper evaluates a 42-learnable-layer network.Source: §2.1 · p. 4
2.2. ResNet
- ResNet uses residual mappings and shortcut connections to address optimization difficulties associated with increased network depth.The study evaluates ResNet-50 and ResNet-152.Source: §2.2 · pp. 4–5
2.3. DenseNet
- DenseNet connects earlier layers’ outputs to later layers within dense blocks, while transition layers compress feature maps.The study evaluates DenseNet-121 and DenseNet-161.Source: §2.3 · pp. 5–6
2.4. CNN Optimization Techniques
- Because the coral datasets are too small to train effective CNNs from scratch, the paper uses transfer learning and data augmentation.The transfer-learning setup initializes from ImageNet weights, replaces the ImageNet classifier with two fully connected layers, and trains those layers.Source: §2.4 · p. 6
3. Previous Advances on Automatic Coral Reef Classification
The section reviews the challenges of underwater coral classification and prior classical and CNN-based approaches. It describes limited dataset portability and the supervision or feature-engineering demands of existing methods.
3.1. Challenges of Coral Classification
- Underwater classification faces occlusion, variable illumination, annotation subjectivity, viewpoint and image-quality variation, within-class variability, class overlap, and motion blur.Available datasets are few and often contain texture patches without global coral structure.Source: §3.1 · p. 7
- Existing datasets often lack whole-coral structure, limiting the input to texture patches.The review also notes few underwater coral datasets and imaging challenges including motion blur.Source: §3.1 · p. 7
3.2. Coral Classification Based on Classical Methods
- Classical approaches commonly combine feature extraction with classification and report results on a single dataset; portability to new datasets has not been demonstrated.Shihavuddin et al. (2013) developed a unified approach for four datasets, including EILAT and RSMAS, using configurable processing steps.Source: §3.2 · pp. 7–8
- The EILAT and RSMAS state-of-the-art classical configuration requires substantial testing and repeated processing, including six algorithms to build a classifier and four for each new image.The configuration combines enhancement, feature extraction, kernel mapping, dimension reduction, prior settings, and classification.Source: §3.2 · p. 8
3.3. Coral Classification Based on CNNs Methods
- CNN-based coral classification can use images without image enhancement or feature extraction, avoiding experiments to find effective combinations of those algorithms.This contrasts with the multi-step classical pipeline described for EILAT and RSMAS.Source: §3.3 · p. 8, §3.2 · p. 8
- Earlier coral CNN studies used VGGnet or LeNet and did not evaluate EILAT or RSMAS; some used hand-crafted features alongside CNNs.Mahmood et al. (2016b) reported 84.5% accuracy on MLC using CNN and hand-crafted features.Source: §3.3 · p. 9
4. Datasets
The study uses two RGB coral texture datasets, EILAT and RSMAS, which differ in image size, collection conditions, class count, and labels. Both have imbalanced class distributions.
4. Datasets
- EILAT and RSMAS are RGB datasets of coral patches that capture texture rather than global coral structure, and both are imbalanced.The datasets are described as among the most recent RGB datasets with the highest number of corals.Source: §4 · pp. 9–10
- The datasets section’s comparison table covers EILAT, RSMAS, and StructureRSMAS, with #imgs denoting the number of images in each class.Source: Table 1 · p. 10
- EILAT contains 1,123 patches sized 64 × 64, collected under equal conditions with one camera and assigned to eight classes.Its labels do not correspond to coral species names.Source: §4 · p. 10
- RSMAS contains 766 patches sized 256×256, collected by divers using different cameras and locations, and assigned to 14 Latin-named coral species classes.Its images were taken under different conditions.Source: §4 · p. 10
5. Classification of Coral Texture Images with CNNs
The section evaluates CNN architectures and training choices for coral texture classification across EILAT and RSMAS, then examines the effects of augmentation and the models’ errors. ResNet models outperform the classical comparison method, while augmentation adds little and some classes remain difficult to distinguish.
5.1. Experimental Framework
- The experiments use five-fold cross-validation and mean fold accuracy, with CNNs initialized from ImageNet weights and the final layers trained for coral classification.The study also varies augmentation settings and network hyperparameters, including iterations, batch size, and depth.Source: §5.1 · p. 13
5.2. Classification of Coral Texture Images without Data Augmentation
- ResNet-152 improves accuracy over Shihavuddin’s method by more than 5% on RSMAS; ResNet-50 and ResNet-152 tie on EILAT and exceed it by more than 2%.Table 2 reports the architecture comparisons without data augmentation.Source: Table 2 · p. 14, §5.2 · p. 15
5.3. Classification of Coral Texture Images with Data Augmentation
- The best data-augmentation configurations improve accuracy by less than 1%, and the authors conclude that augmentation does not significantly improve these coral texture models.The authors attribute the small gains to close-up, small images and transformations that add little variation or lose image information.Source: §5.3 · p. 16
5.4. Analyzing the Missclassified Images
- On EILAT, 18 of ResNet-50’s 22 errors involve Dead Coral, while the remaining three described errors are confusions between Branches Type II and III.The authors note possible bias from EILAT’s class imbalance and report visual similarities between the confused classes.Source: §5.4 · p. 16, Figure 7 · p. 17, Figure 8 · p. 17
- On RSMAS, ResNet-152 makes 10 errors, chiefly confusing APAL with ACER and MCAV with MMEA; the authors suggest expert review and whole-coral images as possible aids.Other RSMAS errors involve blurry images.Source: Figure 10 · p. 18, §5.4 · p. 19
6. Conclusions
CNN models, especially ResNet, achieved state-of-the-art accuracy on EILAT and RSMAS without data augmentation, outperforming classical methods requiring substantial human intervention. The authors report little augmentation benefit for these close-up datasets and identify coral structure images as a future challenge.
6. Conclusions
- Underwater coral classification faces many species, substantial within-species image variation, water-column lighting changes, and overlapping classes when species appear together.Source: §6 · p. 19
- ResNet achieved the best CNN results on EILAT and RSMAS, and CNN models surpassed classical methods requiring high human intervention without data augmentation.The authors report state-of-the-art accuracies on both datasets.Source: §6 · p. 19
- The close-up images in EILAT and RSMAS, taken under similar conditions, showed little benefit from data augmentation.The authors also note substantial within-class variation in these datasets.Source: §6 · p. 19
- The authors identify classifying coral structure images, including with a single classifier for texture or structure images, as a future challenge.Source: §6 · p. 19