Arthroscopic surgery is a major technique for the treatment of joint-related diseases, however, intraoperative bleeding often produces a blood mist that severely affects the surgeon’s field of vision and requires prompt high-flow drainage to remove the mist. Therefore, accurate bleeding detection is a prerequisite for effective blood mist removal. This paper proposes an arthroscopic image bleeding detection method based on the ViT-ResNet50 integrated model and transfer learning to solve the problem of relying on naked eye to identify bleeding in existing arthroscopic surgery. Firstly, Vision Transformer model and ResNet50 model are used to learn features by transfer learning on ImageNet dataset respectively. Then, a difference-enhanced proportional sampling method is proposed to enhance the unbalanced data. Finally, the two sub-network models are integrated through weighted soft voting method to realize bleeding detection in arthroscopic images. In order to evaluate the performance of the model proposed in this paper, experimental results on real data show that the integrated model is superior to a single deep learning model in various performance indicators and has good effects in detecting bleeding in arthroscopic images.

Visual inspection detects manual assembly actions to improve quality. Existing methods mainly analyze behaviors in single assembly scenarios using action recognition algorithms, but they cannot adapt to dynamic product switching and varying station environments in manual assembly lines. This paper proposes an adaptive assembly inspection method based on hand motion trajectories. By constructing an assembly inspection model based on the "Component-Tool-Product" triplet, the YOLOv9 model is employed to capture the spatiotemporal hand trajectory data stream. The assembly process is divided into a sequence of transitions between different assembly regions based on the hand's trajectory. The system analyzes the real-time hand trajectories using Dynamic Time Warping (DTW) and dwell time algorithms to obtain real-time assembly flow, allowing for the detection of missed and misassembled components. Additionally, the frame difference method is introduced to extract hand motion features, and an Attention Feature Fusion (AFF) module is used to integrate multi-scale features with motion texture information, enhancing the performance of the object detection model. Experimental results show that the proposed algorithm effectively reduces the false detection rate, with an average detection accuracy of 96% for missed and misassembled behaviors. The improved YOLOv9 model achieves an mAP@0.5 of 93.94%.