Anomalous events detection in real-world video scenes is a challenging problem owing to the complexity of anomaly and the untidy backgrounds and objects in the scenes. Although there are already many studies on dealing with this problem using deep neural networks, very little literature aims for real-time detection of the anomalous behavior of fish. This paper presents an underwater fish anomalous behavior detection method by combining deep learning object detection, DCG (Directed Cycle Graph), fish tracking, and DTW (Dynamic Time Warping). The method is useful for detecting the biological anomalous behavior of underwater fish in advance so that early countermeasures can be planned and executed. Also, through post-analysis it is possible to access the cause of diseases or death, so as to reduce unnecessary loss, facilitate precision breeding selection, and achieve ecological conservation education as well. A smart aquaculture system incorporating the proposed method and IoT sensors allows extensive data collection during the system operation in various farming fields, thus enabling to develop optimal culturing conditions, both are particularly useful for researchers and the aquaculture industry.

Modern life is generally stressful, so many people choose to keep pets as a source of emotional support. Dogs are one of the most popular choices for pets. However, dogs cannot speak and cannot express their emotions and state of mind. Therefore, being able to effectively interpret a dog’s emotions can help establish and improve the relationship between humans and pets. This paper proposes a system based on deep learning for automatic recognition of dog expressions. The system aims to quickly identify the emotions of pets, thereby predicting the current emotional state of their dogs at any time, and strengthening the connection between owners and their pets. The study uses a dataset of Border Collies containing five emotions for training. The prediction process first uses the YOLOv8 object detection model to detect parts of the face such as ears, eyes, mouth, and nose in single-frame images containing dog faces, and automatically classifies the emotions of these parts. The confidence-weighted method is then used to calculate the highest-scoring emotion in the single-frame image as the decision result for each target. The experiment combines data augmentation and transfer learning to further improve accuracy and obtain better emotion recognition results. Through the proposed method, the study achieves an accuracy of 90%, which is a 32% improvement compared to other related studies abroad.