A multiscale real-time instance segmentation method based oncleaning rubber ball images

Abstract The identification of wear rubber balls in the rubber ball cleaning device in heat exchange equipment directly affects the descaling efficiency. For the problem that the rubber ball image contains impurities and bubbles and the segmentation is low in real time, a multi-scale feature fusion real-time instance segmentation model based on the attention mechanism is proposed for the object segmentation of the rubber ball images. First, we introduce Pyramid Vision Transformer instead of convolution module in the backbone network, and use Spatial-reduction attention layer of transformer to enhance the feature extraction capability at different scales as well as spatial reduction to reduce the computational cost. Then improve the feature fusion module to fuse image features across scales, combined with an attention mechanism to enhance the output feature representation; secondly the prediction head separates the mask branches separately combined with dynamic convolution and depths to improve the accuracy of the mask coefficients. Through the validation of the produced rubber ball dataset, the Dice score, Jaccard coefficient, and mAP of the actual segmented region of this network with the glueball dataset improved by 4.4\%, 4.5\%, and 7.56\%, respectively, and the segmentation speed is also 35.1 fps. The proposed modules can work together to better handle object details and realize better segmentation performance.