Optimzied resnet model of convolutional neural network for under sea water object detection and classification

Abstract

The world’s ocean depths conceal a big mystery, and obtaining the information contained therein is a significant challenge that must be overcome. With the advent of computer vision technologies and robotics, the underwater environment is explored recently. The vast data collected from numerous underwater sensors have a variety of complications related to inadequate image quality, difficulty in acquiring training samples, and uncontrolled objects in the underwater environment. When these images are processed using machine learning techniques that involve manual intervention, the time taken to process a huge amount of images will be relatively high and prone to errors. To tackle these, we propose a novel hybrid capuchin-based coevolving particle swarm optimization (HC²PSO) algorithm with a ResNet model of Convolutional Neural Network (CNN) architecture for underwater object identification. This work mainly aims to explore different underwater objects such as fish, corals, sea urchins, etc. The speckle-reducing anisotropic diffusion (SRAD) filter performs the pre-processing step. The denoising autoencoder (DA) is used for feature extraction which can enhance the partially distorted sample images and offer increased robustness. To overcome the overfitting issue in CNN, the HC²PSO algorithm is used. The experimental works are handled in MATLAB software. Both with and without pre-processing results in terms of SRAD filter are checked and evaluated. The proposed method’s effectiveness is evaluated through various measures like accuracy, specificity, sensitivity, false-positive rate, false-negative rates, etc. The accuracy of the HC²PSO-CNN classifier is higher when compared to the standard CNN classifier in recognizing the underwater objects when evaluated with different performance metrics.

Appendix

The probability density functions can be given as abs(N(0, 1)) and can be determined as,

$$ r(a)=\frac{2}{\sqrt{2\Pi}}\exp -\frac{a^2}{2},\kern0.36em a\ge 0 $$

(26)

The average value of abs(N(0, 1)) is estimated as,

$$ M(a)=\frac{2}{\sqrt{2\Pi}}\int_0^{\infty }{ae}^{-\frac{a^2}{2} da}=0.798 $$

(27)

The variance value can be calculated for the abs(N(0, 1)) as shown below,

$$ V(a)={\sigma}^2\frac{2}{\sqrt{2\Pi}}\int_0^{\infty }{\left(a-\varsigma \right)}^2{e}^{\frac{-{a}^2}{2}} da $$

(28)

$$ {\sigma}^2=\frac{2}{\sqrt{2\Pi}}\int_0^{\infty }{\left(a-\varsigma \right)}^2{e}^{\frac{-{a}^2}{2}} da=0.36 $$

(29)

From the above eqn, the obtained SD is σ = 0.60. Then it is essential to produce the stochastic coefficients for the terms (PS_i − a_i(t)) and (PS_η − a_i(t)). Henceforth the velocity equation can be upgraded as,

$$ {V}_i\left(t+1\right)=\left| rann\right|\left({PS}_i-{a}_i(t)\right)+\left|\mathit{\operatorname{ran}}1n\right|\left({PS}_{\eta }-{a}_i(t)\right) $$

(30)

The positive random number generated by the abs(N(0, 1)) is represented as |rann| and |ran1n|. Assume a random variable c that has been obtained from the sum of the above-mentioned random variables. The probability density function is considered as p(a). from this, the probability distribution function is determined as,

$$ p(a)=\frac{2}{\sqrt{\Pi}}{e}^{-\frac{c^2}{4}}.\mathit{\operatorname{erf}}(0.5c) $$

(31)

Here the value of erf(c) is defined as,

$$ \mathit{\operatorname{erf}}(c)=\left(1/\sqrt{\Pi}\int_0^c{e}^{-{a}^2} da\right) $$

(32)

Thus the generation of stochastic coefficients with the aid of abs(N(0, 1)) enables the compromise among a large number of low amplitudes and the less number of maximum amplitudes i.e., fine-tuning of parameters.