Search

Deep learning-based computer vision technology has been extensively utilized to assist medical microscopic smear diagnoses, such as complete blood count, detecting mycobacterium tuberculosis in sputum samples, and identifying plasmodium in blood smears. Effective medical microscopic smear detection necessitates balancing both detection accuracy and model parameters, facilitating the application of the model on low-resource computing platforms and ensuring real-time detection capabilities. Furthermore, the model should demonstrate superior generalization capabilities on microscopic smears to meet various detection tasks. In this study, we propose YOLO-FMS, a lightweight and efficient model based on YOLOv5 for medical microscopic smear detection with a compact weight of 15.06M, which can resolve the challenge of the balance between detection accuracy and model parameters. Firstly, YOLO-FMS improves the performance of small-scale platelet and mycobacterium tuberculosis detection by adding a small target detection head. Secondly, A lightweight convolutional technique, GSConv, was introduced to make the symbolic ability of the lightweight convolutional as close to the vanilla convolutional as possible and to reduce the computational cost. Thirdly, the feature extraction ability of YOLO-FMS is enhanced by C3-B-CBAM and Tiny-SPPCSPC modules constructed by our proposed method. Comprehensive experiments prove that YOLO-FMS shows high detection accuracy, with 92.5% mAP on the BCCD dataset and 87.6% mAP on the Tuberculosis-Phonecamera dataset. Additionally, numerous verification experiments are conducted on the BCDD and BBBC041 datasets, and the results confirm the effectiveness and generalization capability of YOLO-FMS in the medical microscopic smear detection field. The codes of YOLO-FMS are available at https://github.com/GefionP/YOLO-FMS.

Synthetic Aperture Radar (SAR) imaging technology is crucial for maritime vessel monitoring. However, the inherent characteristics of SAR images, such as limited feature resolution and speckle noise, pose a series of challenges for ship detection. Although existing research has made great progress in improving the model detection accuracy, it often comes at the cost of reduced efficiency and increased computational complexity. To address this issue, we propose a lightweight SAR ship detection model based on YOLOv5s, named GS-YOLO. Firstly, we adopt C3GhostV2 modules as the lightweight backbone to enhance the computational efficiency which is based on GhostNetV2 and GConv. In the neck, Squeeze and Excitation (SE) attention is employed to strengthen the feature extraction ability for small targets and improve the SAR ship detection precision. Furthermore, we develop a novel XIoU loss function to reinforce the accuracy and robustness of our model. Comparative experiment results on the HRSID dataset show that the proposed GS-YOLO achieves a significant increase in detection precision from 88.2% to 92.7% and the mean Average Precision (mAP) from 90.5% to 94.3%. In addition, the parameters are effectively reduced, achieving a good balance between detection speed and accuracy.

This paper explores the scientific & technological innovation performance of world-class universities in China from 2014 to 2019, based on the super-efficiency DEA model and Malmquist index. Then we provide a novel five-dimensional analysis framework to research influencing factors of their S&T innovation efficiency with the Tobit-DEA model. The results show that 36.6% of the sample universities are DEA efficient, and their efficiency varies according to their type and location. Besides, the technical efficiency index of universities is relatively high and the total factor productivity is volatile, which are reflected in the ability of transforming scientific research achievements into practical productivity. From the perspective of dynamic analysis, the sample universities expand slowly in the frontier of S&T innovation efficiency. In addition, the S&T innovation performance is highly related to the input quality and matching structure of scientific research elements, government relevance, and industry-academia-research collaboration level. Among these main factors, the high proportion of full-time teachers with senior titles, reasonable resource allocation structure and government support have significant positive impact, whereas the number of participants in international academic has negative impact.

In recent years, tactile sensors have experienced tremendous progress. Among them, piezoresistive tactile sensors using nanocomposites exhibit good flexibility and sensitivity. Due to their low cost and good performance, piezoresistive tactile sensors have great potential for large-scale commercialization. However, the drawbacks are equally obvious, i.e., poor consistency and low-sensing range. The fundamental reason is that conductive fillers have poor decentralization in polymer and lacks an effective structural design. To solve these problems, we compared six common organic solvents in the solution blending method to prepare the multi-walled carbon nanotubes/polydimethylsiloxane (MWCNT/PDMS) composite. The optimal organic solvent by comparing the dispersion performance of MWCNT and the stability of MWCNT/PDMS is presented. Moreover, the well-consistent tactile sensors were proposed and fabricated using the micro-electromechanical systems technologies. The sensitivity reached 6.25% KPa-1, and the reproducibility test of the same sensor has been performed. The result proves that the proposed schemes for nano-materials preparation and processing have greatly improved the consistency of the piezoresistive tactile sensors.

Automatic grasping is playing an important role in robotics, and traditional grasping approaches cannot deal with both occlusions and texture-less objects well. To improve the stability and accuracy of grasping in a novel and occlusion environment, a grasping approach based on instance segmentation and self-supervised learning pose estimation network is proposed to grasp objects with the manipulator in this paper. The approach can be divided into three phases: instance segmentation, pose estimation, and pose transformation. Instance segmentation predicts classification masks for each pixel. Masks which stand for the contour of objects provide a heuristic knowledge for the self-supervised learning pose estimation network. Pose estimation network has two fully connected layers and regards estimation as a self-supervised classification problem. Then, the pose can be transformed from value in pixel coordinates to actual value relative to the base coordinate of the manipulator. As a result, the manipulator can be operated to grasp by the given actual value of pose. With the help of the approach proposed in this paper, we improve the grasping accuracy by 35%, compared to the former grasping approach based on the pose estimation network on the grasping dataset of CMU. Besides, grasping with this approach on hardware also shows a high success rate. Therefore, the proposed approach is a more robust and more accurate way of grasping.