This study proposes Combining Attention and Brightness Adjustment Network (CABA-Net), a deep learning network for underwater image restoration, to address the issues of underwater image color-cast, low brightness, and low contrast. The proposed approach achieves a multi-branch ambient light estimation by extracting the features of different levels of underwater images to achieve accurate estimates of the ambient light. Additionally, an encoder-decoder transmission map estimation module is designed to combine spatial attention structures that can extract the different layers of underwater images’ spatial features to achieve accurate transmission map estimates. Then, the transmission map and precisely predicted ambient light were included in the underwater image formation model to achieve a preliminary restoration of underwater images. HSV brightness adjustment was conducted by combining the channel and spatial attention to the initial underwater image to complete the final underwater image restoration. Experimental results on the Underwater Image Enhancement Benchmark (UIEB) and Real-world Underwater Image Enhancement (RUIE) datasets show excellent performance of the proposed method in subjective comparisons and objective assessments. Furthermore, several ablation studies are conducted to understand the effect of each network component and prove the effectiveness of the suggested approach.
In aquaculture, the presence of dead fish on the water surface can serve as a bioindicator of health issues or environmental stressors. To enhance the precision of detecting dead fish floating on the water’s surface, this paper proposes a detection approach that integrates data-driven insights with advanced modeling techniques. Firstly, to reduce the influence of aquatic disturbances and branches during the identification process, prior information, such as branches and ripples, is annotated in the dataset to guide the model to better learn the scale and shape characteristics of dead fish, reduce the interference of branch ripples on detection, and thus improve the accuracy of target identification. Secondly, leveraging the foundational YOLOv8 architecture, a DD-IYOLOv8 (Data-Driven Improved YOLOv8) dead fish detection model is designed. Considering the significant changes in the scale of dead fish at different distances, DySnakeConv (Dynamic Snake Convolution) is introduced into the neck network detection head to adaptively adjust the receptive field, thereby improving the network’s capability to capture features. Additionally, a layer for detecting minor objects has been added, and the detection head of YOLOv8 has been modified to 4, allowing the network to better focus on small targets and occluded dead fish, which improves detection performance. Furthermore, the model incorporates a HAM (Hybrid Attention Mechanism) in the later stages of the backbone network to refine global feature extraction, sharpening the model’s focus on dead fish targets and further enhancing detection accuracy. The experimental results showed that the accuracy of DD-IYOLOv8 in detecting dead fish reached 92.8%, the recall rate reached 89.4%, the AP reached 91.7%, and the F1 value reached 91.0%. This study can achieve precise identification of dead fish, which will help promote the research of automatic pond patrol machine ships.
Underwater fish image segmentation is the key technology to realizing intelligent fisheries and ecological monitoring. However, the problems of light attenuation, blurred boundaries, and low contrast caused by complex underwater environments seriously restrict the segmentation accuracy. In this paper, RetinalConet, an underwater fish segmentation network based on bionic retina dual-channel and multi-module cooperation, is proposed. Firstly, the bionic retina dual-channel module is embedded in the encoder to simulate the separation and processing mechanism of light and dark signals by biological vision systems and enhance the feature extraction ability of fuzzy target contours and translucent tissues. Secondly, the dynamic prompt module is introduced, and the response of key features is enhanced by inputting adaptive prompt templates to suppress the noise interference of water bodies. Finally, the edge prior guidance mechanism is integrated into the decoder, and low-contrast boundary features are dynamically enhanced by conditional normalization. The experimental results show that RetinalCoNet is superior to other mainstream segmentation models in the key indicators of mDice, reaching 82.3%, and mIou, reaching 89.2%, and it is outstanding in boundary segmentation in many different scenes. This study achieves accurate fish segmentation in complex underwater environments and contributes to underwater ecological monitoring.