Wu-Chih Hu

Color cast is a novel challenge in underwater image enhancement (UIE) tasks. In this study, an innovative enhancement method, named AIUIE-Net, that is specifically designed to address the color cast problems that are commonly found in underwater images is proposed. AIUIE-Net integrates two key components: a water depth-based color cast correction and channel compensation network (WDC-Net) that uses quasilinear water depth parameters for adaptive color cast correction and compensation and an optical particle-based high-fidelity restoration network (OP-HFRNet) that is a physics-informed deep learning backbone that simulates the propagation of light particles in underwater environments. WDC-Net leverages water depth information to perform precise color correction, and OP-HFRNet enhances the realism of underwater scenes by modeling the physical behavior of optical particles. To the best of our knowledge, AIUIE-Net is the first UIE framework to integrate color cast correction and optical particle-based high-fidelity restoration. Experimental results demonstrate that AIUIE-Net outperforms existing state-of-the-art (SOTA) methods across multiple evaluation metrics. It produces enhanced images with high fidelity, clarity, vividness, and saturation, thus effectively restoring the stunning authenticity and visual quality of the underwater world.

Classifying orchids is a challenging task: the color, shape and texture features of each part of an orchid image are small, the category similarity is high, and the features are blurred. Traditional deep learning methods are not suitable for orchid image classification, mainly because they rely on a large amount of human interpretation for labeling models to achieve a certain degree of accuracy. Therefore, to address the human interpretation problem of orchid classification, we propose a selfsupervised learning-based efficient multiscale feature fusion convolutional neural lightweight network architecture called OrchidNet. OrchidNet is divided into three parts: a convolutional neural network (CNN) attention network, a self-supervised finegrained mechanism, and a feature pyramid network for multiscale feature fusion. In this paper, we focus on lightweight and highperformance materials. The motivation for our method is to use convolutional neural networks to upsample the backbone architecture (backbone) of local visual features. In orchid images, most of the features are extracted by the CNN backbone (attention mechanism). Afterward, the self-supervised fine-grained mechanism is used for feature screening. Finally, the global and local feature information is downsampled through the feature pyramid for multiscale feature fusion. This study significantly impacts the image-processing community of orchid classification. The proposed OrchidNet improves the accuracy of orchid classification and uses a self-supervised learning method that does not require manual labeling, thereby reducing the labeling time and the cost of manual interpretation. The experimental results show that the proposed OrchidNet is superior to state-of-the-art methods for orchid classification used in recent years.

The fish industry is an important source of income for island countries. Fish is a main source of animal-based protein. Marine fishing is gradually being replaced by marine farming (or aquaculture) due to declining wild fish populations and water pollution. However, fish farming is costly job with high requirements for labor, electricity, water, and feed. The use of deep learning to perform intelligent surveillance in aquaculture fields, reducing the need for human resources and implementing real-time monitoring, has been proposed. In this article, we propose a novel deep residual network (ResNeXt 3 $\times$ 1 $D$ ) for abnormal behavior recognition in aquaculture fields. The proposed ResNeXt 3 $\times$ 1 $D$ convolutional network is mainly based on an R(2 $+$ 1) $D$ convolutional network and modified to obtain better performance. Experimental results showed that the proposed ResNeXt 3 $\times$ 1 $D$ exhibited good performance for abnormal behavior recognition in aquaculture fields. Specifically, the accuracy obtained using the proposed ResNeXt 3 $\times$ 1 $D$ for abnormal behavior recognition in aquaculture fields was approximately 95.3%.

The Artificial Intelligence of Things (AIoT) plays a crucial role in shrimp farming by enabling automated and real-time monitoring of shrimp counting, especially larvae. With this counting information, proper feeding control can be maintained by equipping IoT devices to capture real-time data on water quality, temperature, and other environmental factors, ensuring healthy shrimp growth and increasing production. Taking advantage of the power of AIoT, this paper proposes a Scaled MutI-Layer fEature fuSion Network (SMILES-Net) to allow farmers to remotely and automatically manage the counting of shrimps, effectively reducing the need for manual labor, and enabling swift interventions to sustain shrimp well-being. Since shrimp larvae are extremely small, we frame this counting problem as a density prediction problem, where the sum of the constructed density map is the total number of shrimps predicted. The pooling operation used in convolutional neural networks scales each feature map to 1/4 and causes the rich features of small shrimps to disappear dramatically. To tackle this truncation problem, we propose a novel Synthetic Fusion Module (SFM) and an Intra-block Fusion Module (IFM) to create a smoother scale space, generating better heat maps with fine-grained features for shrimp counting. Furthermore, we introduce a lightweight version of SMILES-Net (LW-SMILES-Net) that enables real-time shrimp counting without compromising accuracy. This method is evaluated on different datasets for shrimp counting and outperforms all SoTA methods. Overall, integrating SMILES-Net with IoT devices can provide a powerful solution for real-time shrimp larvae counting in shrimp farming, contributing to sustainable increases in shrimp production. The dataset is available at https://github.com/Naughty725/shrimp.

In aquaculture, controlling water quality parameters is an important challenge. The water quality parameters affect the growth of aquatic organisms. Thus, maintaining water quality balance has become the primary goal of aquaculture operators. However, the traditional water quality inspection method is low in accuracy and consumes considerable time and human resources. On the other hand, since water quality sensors are immersed in seawater for a long time, algae will grow on the sensors, affecting their accuracy. Therefore, to solve the abovementioned problems, this paper reports the design and implementation of a full-time artificial intelligence of things (AIoT)-based water quality inspection and prediction system, which uses a simple recurrent unit (SRU) model to predict water quality data. With the proposed system, it is possible to collect water quality sensing data 24 hours a day and further use the SRU model for sensor data prediction to assist aquaculture farmers in managing and controlling outdoor aquaculture ponds. Moreover, a 24-hour water quality sampling tank is designed to overcome the problem of sensor error. Throughout the whole process, data are transmitted to a water quality monitoring cloud platform for further inspection and prediction. In this paper, SRU-based prediction is used to obtain predictions of water quality parameters, and three popular metrics (MAE, RMSE, and MAPE) are used to evaluate the performance. As a result, experimental results show that the proposed method offers good performance for prediction of water quality.

In the process of white shrimp farming, shrimp larvae trading is a time-consuming and laborious process. In the traditional shrimp larvae counting process, it is often necessary to count the number of shrimp larvae for sale through manual counting. However, there is not a unified standard method for counting. Computer counting methods can only identify a small number of non-overlapping shrimp larvae, and the accuracy of the methods is not high. In recent years, with the advent of deep learning, there have been many applications of counting objects. It is desirable to develop a method for counting shrimp larvae that is less labor-intensive and can achieve fast results. In this paper, a deep learning-based fast counting methodology using density estimation for counting shrimp larvae is proposed, with noteworthy advantages over traditional shrimp larvae counting methods. The experimental results showed that the average accuracy of the proposed methodology was as high as 98.72% when the number of shrimps was approximately 1,000. Moreover, the average execution time of the proposed methodology was around 12.35 s, so that the average execution time can be accepted in the real world.

The decisions made regarding traditional fish feeding systems mainly depend on experience and simple time control. Most previous works have focused on image-based analysis of the leftover feed at the bottom of the pond to determine whether to continue or to stop feeding. However, the feasibility of such a method in an actual outdoor aquaculture pond is low. The main reason for this is that real outdoor aquaculture ponds have turbid water quality, small feed targets, interference from intense fish activity, overlapping images of fish and feed, etc. Therefore, image-based recognition is not easy to implement in actual outdoor aquaculture. To overcome this problem, this article proposes an automatic fish feeding system based on deep learning computer vision technology. In contrast to traditional computer-vision-based systems for recognizing fish feed underwater, the proposed system uses deep learning technology to recognize the size of the waves caused by fish eating feed to determine whether to continue or to stop feeding. Furthermore, several water quality sensors are adopted to assist in feeding decisions. As a result, the proposed system uses deep learning technology to recognize the size of the water waves caused by fish eating feed to determine whether to continue to cast feed or to stop feeding. Experimental results show that an accuracy of up to 93.2% can be achieved.