PaperlayerView exact revision on arXiv

Source-linked AI summary

Underwater Fish Species Classification using Convolutional Neural Network and Deep Learning

Dhruv Rathi, Sushant Jain, Dr. S. Indu

arXiv:1805.10106v1cs.CV

TL;DR

The paper addresses the difficulty of classifying fish in underwater imagery, where noise, distortion, overlap, and occlusion complicate classification. It combines image preprocessing with a CNN and reports 96.29% accuracy.

  • Problem

    Underwater fish classification has seen no significant breakthrough in datasets from underwater videos, amid noise, distortion, overlap, segmentation errors, and occlusion.

    Source: §I · p. 1

  • Method

    The method preprocesses Fish4Knowledge images with image-processing operations and feeds the enhanced images to a convolutional neural network for species classification.

    Source: §I · p. 1, §III · pp. 2–3

  • Results

    96.29% accuracy; computation time was 0.00183 seconds per frame, which the authors say supports real-time applications.

    Source: §V · p. 5

  • Takeaways & Limitations

    The authors conclude that the approach can be used for real-time fish-species classification applications.

    Source: §V · p. 5

    Background noise and other water bodies affected some images, preventing 100% accuracy.

    Source: §V · p. 5

Abstract

from arXiv · show

The target of this paper is to recommend a way for Automated classification of Fish species. A high accuracy fish classification is required for greater understanding of fish behavior in Ichthyology and by marine biologists. Maintaining a ledger of the number of fishes per species and marking the endangered species in large and small water bodies is required by concerned institutions. Majority of available methods focus on classification of fishes outside of water because underwater classification poses challenges such as background noises, distortion of images, the presence of other water bodies in images, image quality and occlusion. This method uses a novel technique based on Convolutional Neural Networks, Deep Learning and Image Processing to achieve an accuracy of 96.29%. This method ensures considerably discrimination accuracy improvements than the previously proposed methods.

I. INTRODUCTION

The introduction motivates fish classification as a tool for ecological monitoring and identifies underwater imagery as a difficult, under-served setting. It proposes preprocessing Fish4Knowledge images and classifying them with a CNN.

I. INTRODUCTION

  • Visual classification can inform ecological monitoring through species counts, distributions, movement traces, and activity patterns.The paper connects these observations to ecosystem health and environmental change.Source: §I · p. 1
  • Underwater fish classification remains challenging because images contain noise, distortion, overlap, segmentation errors, occlusion, shifting colors, inconsistent lighting, sediments, and plants.Prior work has more often classified fish outside water or in adequately lit tanks [2][3][4][5].Source: §I · p. 1
  • The proposed approach preprocesses Fish4Knowledge images with Gaussian blurring, morphological operations, Otsu thresholding, and pyramid mean shifting before CNN classification.The authors present CNNs as simpler, more robust, and adaptable to incoming data as the dataset matures.Source: §I · p. 1

II. RELATED WORKS

The reviewed work uses varied image features, classifiers, and detection strategies, with results reported across datasets of different sizes and species scopes. Several studies note constraints tied to species coverage, image conditions, or computational complexity.

II. RELATED WORKS

  • Texture and shape methods reported 92% accuracy on 360 images, while morphological filters distinguished three species at 90%, 88%, and 90%.The latter method depends on morphological-filter size and shape.Source: §II · pp. 1–2
  • Haar classifiers were sensitive to background and viewing angle, while image warping with SVM reached 90% on 320 images.Source: §II · p. 2
  • Shape-and-texture recognition on 300 test and 600 training images achieved 81.67% with EDM and 99.00% with ANN.Source: §II · p. 2
  • SIFT and PCA reached 92% across six species but used only 162 images, while deformable-shape detection was limited by computational complexity.Source: §II · p. 2
  • Underwater video detection, tracking, and counting reached 85% accuracy across 20 videos, while neural-network identification reached 91.65% on 697 images.Source: §II · p. 2
  • Gabor features yielded 70.6% and 80.3% for two species; other studies reported 72% on wild cichlids, 81.8% at a feature-space dimensionality, and 78.59%.The cichlid study targeted only that fish group.Source: §II · p. 2
  • Infrared contour classification reported around 78%, landmark analysis used only 22 images across nine species, and descriptor selection reported about 85%.The contour method performed poorly with noisy inputs.Source: §II · p. 2

III. PROPOSED METHODOLOGY

The method combines image preprocessing with a CNN to classify fish species. It enhances images through thresholding and morphological operations, then feeds the original and processed images to the network.

III. PROPOSED METHODOLOGY

  • Otsu thresholding creates a fish-focused foreground; erosion thins it, and dilation joins image regions broken during noise removal.The preprocessing is intended to remove underwater obstacles, dirt, and non-fish bodies before CNN training.Source: §III · pp. 2–3
  • The CNN receives a 100×100×4 input formed by stacking the original 100×100×3 RGB image with the 100×100×1 preprocessed image.The network uses convolutional and pooling layers, followed by fully connected layers and an 80% dropout layer.Source: §III · pp. 2–4
  • Convolutional units connect to subsets of prior-layer units with contiguous receptive fields; shared weights support position-independent feature detection and reduce free parameters.Max-pooling downsamples non-overlapping image regions by retaining each region’s maximum value.Source: §III · pp. 3–4, Figure 2a · p. 3, Figure 2b · p. 3
  • 32 5×5×4 filters applied to a 100×100×4 feature map produce a 96×96×32 output, followed by subsampling and further convolutional processing.The described subsequent stages use 64 filters and then 32 filters, with max-pooling; a fully connected layer is followed by 80% dropout and another fully connected classification layer.Source: §III · pp. 3–4, Figure 3 · p. 3
  • The method compares ReLU, sigmoid, and tanh activation functions across CNN layers and uses cross-entropy after the fully connected layer.The cross-entropy is described as a negative log-likelihood for the data under the model.Source: §III · p. 4

IV. EXPERIMENTAL RESULTS

The method was tested on Fish4Knowledge images, and Tables 2 and 3 cover species sample counts and test-image accuracy.

IV. EXPERIMENTAL RESULTS

V. CONCLUSION AND FUTURE WORK

The proposed method reports 96.29% accuracy and a computation time of 0.00183 seconds per frame. The authors identify image noise and other water bodies as error sources and plan image enhancement as future work.

V. CONCLUSION AND FUTURE WORK

  • 96.29% accuracy is reported, described as very high compared with other current methods for this application.Source: §V · p. 5
  • 0.00183 seconds per frame is the reported computation time, and the authors say the method can be used for real-time applications.Source: §V · p. 5
  • Background noise and other water bodies contributed to images that could not be classified accurately, preventing 100% accuracy.Source: §V · p. 5
  • Future work plans to implement image-enhancement techniques to counter lost image features.Source: §V · p. 5
Loading 1805.10106v1…