To address the poor performance of underwater object detection caused by light attenuation and scattering, this paper proposes an enhanced underwater object detection framework based on YOLOv8, named ERMS-YOLOv8, aiming to improve detection accuracy. The backbone is replaced with an efficient vision transformer（EfficientViT）to strengthen feature extraction of underwater organisms and reduce information loss. The neck adopts a reparameterized generalized-directional feature pyramid network（RepGFPN）to enhance the fusion of high-level semantic and low-level spatial features, enabling richer feature representation. A mixed local channel attention for object detection（MLCA）is introduced to integrate channel, spatial, local, and global channel information, thereby boosting the model's representational capacity. Additionally, a scalable intersection over union loss（SIoU）is employed to improve boundary prediction accuracy. Experimental re sults demonstrate that the proposed method achieves mAP values of 83.9% on the UPRC2021 dataset and 84.4% on the DUO dataset, outperforming the original YOLOv8 and exhibiting superior performance in underwater object detection.

A long short-term memory（LSTM）neural network edge computing accelerator based on distributed systolic array architecture was proposed on the resource limited edge computing devices. The design distributes input data storage to reduce data movement and power consumption, while data transmission in a systolic manner minimizes the idle rate of computing units and enhances computational efficiency. Experimental validation on a VU13P field-programmable gate array（FPGA）shows that the proposed LSTM accelerator achieves an effective computing power of 179.2 GOPS at an operating frequency of 200 MHz, with a dynamic power consumption of 0.343 W and an energy efficiency of 522.4 GOPS/W. Compared with typical existing designs, the proposed accelerator improves energy efficiency by more than 34%.

To address the computational bottlenecks faced by classical neural networks under the explosive growth of data scale, quantum convolutional neural networks（QCNNs）based on quantum computing have become a research hotspot. This study constructs a QCNN for image classification within the limited resources provided by noisy intermediate-scale quantum（NISQ）devices. The model employs angle encoding and designs a convolutional layer based on a data re-uploading classifier, followed by a four-qubit pooling layer. Two different architectures of quantum fully connected layers are designed to perform image classification, and the impact of their structures on QCNN classification performance is analyzed.Simulation results show that the proposed QCNN achieves high classification accuracy and good generalization in binary classification tasks, with a maximum accuracy of 100.00%, a minimum of 94.55%, and an average of 97.29%. Furthermore, increasing the circuit depth improves model performance, enabling the QCNN to achieve over 90% accuracy in fourclass classification tasks.

The current dialogue emotion recognition models often overlook the coherence features and discourse structure information in context modeling. Therefore, this paper proposes a dialogue emotion recognition model based on coherence features and discourse structure. Firstly, discourse coherence detection is conducted to eliminate weak or incoherent discourse information, and both local and global coherent information are obtained by constructing a coherence matrix. Secondly, a dialogue parser is utilized to establish discourse structure relations, and a directed acyclic graph is employed to model the discourse structure while conveying both discourse structure information and speaker information. Finally, through interactive attention, coherent information and discourse information are interactively integrated to generate emotional labels. This paper validates the proposed model using two public datasets, with results indicating that compared to existing models, the proposed model demonstrates certain improvements in performance indices.

To alleviate the scarcity of data, this paper collects and annotates two infrared tracking datasets for sequential small target detection, named ATR-ISTD and UAV-ISTD. This paper proposes a sequential small target detection network integrating a memory pool, which effectively utilizes the correlation information between frames before and after, reads memory information through memory matching between the query frame and the memory frame, and solves the problems of high false alarm and low accuracy in infrared small target detection under high clutter background. To reduce the loss of small target features caused by downsampling, a forward semantic guided fusion module（PSGF）is designed to integrate features of different scales. In the memory vector encoder, a pseudo label guided feature enhancement module（PLG-FE）is designed to enhance the local feature expression ability of small targets. Experimental results show that, compared with mainstream single-frame detection methods, the proposed method significantly reduces false alarm rates, achieving improvements of 16.87% and 10.49% on the ATR-ISTD and UAV-ISTD datasets, respectively. Target-level F1 scores increased by 4.89% and 6.54%, and pixel-level F1 scores improved by 7.69% and 11.63%.