Tibetan Jiu chess is a national intangible cultural heritage of People’s Republic of China. The digital game platform has collected some online game data, but due to life habits, most masters and enthusiasts prefer to play offline. To collect offline gameplay data efficiently and quickly for Jiu chess, this paper proposes the intelligent algorithm based on improved You Only Look Once (YOLO) version 8 nano (YOLOv8n) model. Firstly, in order to improve the multi-scale feature fusion ability, the backbone and neck are replaced with co-designing and scaling ConvNets with masked autoencoders (namely ConvNeXt V2) and Gather-and-Distribute (GD) mechanism, respectively. Secondly, to reduce the parameter size of the optimized model, the Layer-Adaptive Magnitude-based Pruning (LAMP) score channel pruning algorithm is proposed. Finally, in order to convert the output of chess piece detection into standard chess manual data, the nearest Euclidean distance information matching algorithm and the game conversion algorithms are also proposed, separately. The experimental data source comes from the video recorded in 2023 National Tibetan Jiu Chess Competition in Qinghai Province. The experimental results show that the accuracy of our proposed model reaches 99.660% compared with other models, and the detection time after pruning is reduced by 5.5 ms. The results confirm that our algorithm can effectively improve and balance the speed and accuracy of chess piece detection.

Tibetan Jiu Chess, a national intangible cultural heritage of China, involves a complex gameplay with both layout and battle stages, presenting a vast action space. Deep learning effectiveness is often limited by computational and data constraints. To address this, we propose LSAFJiu, a low-resource algorithm based on long sequence action forecasting(LSAF). Its innovation combines an LSAF model, policy network, and value network to evaluate game tree leaf nodes. The LSAF model incorporates the ActionFormer for the layout stage and an actionstate autoencoding planner for the battle stage, aimed at enhancing decision-making capabilities while reducing resource consumption. The policy and value networks are built on the ResNet50 architecture, utilizing a serial stack activation function to minimize hardware requirements and improve prediction accuracy. To train the initial policy network, 20,000 high-quality game instances were generated by three intelligent programs grounded in human expertise, effectively reducing self-play training time. Experiments conducted on two 2060 GPUs demonstrated that LSAFJiu increased the game’s winning rate by 30.5% while requiring only 2.51% of the simulation time needed by the policy network alone. Additionally, the serial stack activation function improved prediction accuracy by at least 2%. The overall winning rate of LSAFJiu reached 88% within a reasonable decision time. LSAFJiu is an efficient, low-resource algorithm that offers accurate decision-making for Tibetan Jiu Chess.

Few-shot object detection receives much attention with the ability to detect novel class objects using limited annotated data. The transfer learning-based solution becomes popular due to its simple training with good accuracy, however, it is still challenging to enrich the feature diversity during the training process. And fine-grained features are also insufficient for novel class detection. To deal with the problems, this paper proposes a novel few-shot object detection method based on dual-domain feature fusion and patch-level attention. Upon original base domain, an elementary domain with more category-agnostic features is superposed to construct a two-stream backbone, which benefits to enrich the feature diversity. To better integrate various features, a dual-domain feature fusion is designed, where the feature pairs with the same size are complementarily fused to extract more discriminative features. Moreover, a patch-wise feature refinement termed as patch-level attention is presented to mine internal relations among the patches, which enhances the adaptability to novel classes. In addition, a weighted classification loss is given to assist the fine-tuning of the classifier by combining extra features from FPN of the base training model. In this way, the few-shot detection quality to novel class objects is improved. Experiments on PASCAL VOC and MS COCO datasets verify the effectiveness of the method.