Improved YOLOv3 and U-Net Based Pipetting Abnormal Condition Identification and Pipetting Region Segmentation

To improve the reliability of micro pipetting in automatic pipetting workstations, this paper analyzes and investigates the problems related to abnormal condition recognition and pipetting volume detection during pipetting using deep learning models. Firstly, the YOLOv3 network model is designed to be lightweight, and the DarkNet-53 network, which is the backbone of the network structure for image feature information extraction, is replaced by the MobileNetv3 network to realize the recognition of abnormal conditions such as empty aspiration, air bubbles and blockage during pipetting; then the U-Net network model is built to integrate the CBAM attention mechanism module and the residual module to realize the segmentation of pipetting areas. Then, we construct a U-Net network model incorporating the CBAM attention mechanism module and residual module to segment the pipetting area. The results showed that the model size of the improved YOLOv3 network was reduced by 139.1 MB and the frame rate per second (FPS) was increased by 15.7 frames/s. The improved U-Net network improved the IOU and F1 values for pipetting area segmentation by 8% and 13% respectively.