Accurate Fish Detection under Marine Background Noise Based on the Retinex Enhancement Algorithm and CNN

Abstract

Underwater detection equipment with fish detection technology has broad application prospects in marine fishery resources exploration and conservation. In this paper, we establish a multi-scale retinex enhancement algorithm and a multi-scale feature-based fish detection model to improve underwater detection accuracy and ensure real-time performance. During image preprocessing, the enhancement algorithm combines the bionic structure of the fish retina and classical retinex theory to filter out underwater environmental noise. The detection model focuses on improving the detection performance on small-size targets using a deep learning method based on a convolutional neural network. We compare our method to current mainstream detection models (Faster R-CNN, RetinaNet, YOLO, SSDetc.), and the proposed model achieves better performance, with a mean Average Precision (mAP) of 78.31% and a mean Miss Rate (mMR) of 54.11% in the open fish image data set. The test results for the data from the field experiment prove the feasibility and stability of our model.

An Underwater Dense Small Object Detection Model Based on YOLOv5-CFDSDSE

“…How-ever, the detection results are unstable due to the extreme complexity of the underwater optical environment. Chen et al [14] proposed a multiscale Retinex enhancement algorithm that combined the Retinex algorithm by emulating the fish retina to eliminate underwater noise. They also used deep learning methods to improve the detection performance of small objects.…”

High Speed and Precision Underwater Biological Detection Based on the Improved YOLOV4-Tiny Algorithm

“…YOLOV4-Tiny [25] is based on YOLOV4, which uses the most up-to-date network structure and training skills to achieve high accuracy and speed. The main improvements are shown below: (1) The backbone network is CSPDarknet53-Tiny instead of CSPDarknet53, and only 13 × 13 and 26 × 26 scale feature layers are used in the model structure. The CSPDarknet53-Tiny consists of a stacked CBL (Convolution, Batch Normalization, and Leaky-ReLU) structure and a CSPrestblock_body (Cross Stage Partial restblock body) structure [26].…”

“…Recently, the application value of underwater detection technology has attracted a lot of attention. Optical sensing is a critical information acquisition source for underwater detection equipment due to its rich and intuitive perception information [1]. At present, underwater object detection based on optical images has many applications in the marine environment, including the study of marine ecosystems, marine biological population estimation, marine species conservation, pelagic fishery, and underwater unexploded ordnance detection [2].…”

“…This paper provides a lightweight detection network and an underwater data augmentation algorithm. The main contributions of this paper are as follows: (1) The CBAM self-attention mechanism and the Ghost module are used to improve the YOLOV4-Tiny. In this way, the useful underwater object feature information is emphasized and effectively minimizes the running time, while retaining excellent detection precision.…”