Semantic segmentation of mechanical assembly using selective kernel convolution UNet with fully connected conditional random field

Monitoring a mechanical assembly is vital for ensuring the quality of the mechanical products. In this study, each part of a mechanical assembly is recognized via precise segmentation of the mechanical assembly images to determine the assembly sequence of the mechanical products as well as to detect missing and false assemblies. For the segmentation of components in mechanical assembly images, this study proposes a method that combines a selective kernel convolution UNet with a fully connected conditional random field (DenseCRF) (SKC-UNet + DenseCRF). In the proposed SKC-UNet, an improved SKC-Net block is introduced in the coding network of UNet, which enables the neurons to automatically adjust the size of the receptive field on the basis of multiple scales of the received information. Thus, a dynamic selection mechanism can be realized and the number of parameters can be reduced drastically; therefore, the network becomes simpler. DenseCRF provides an image data-dependent smoothing term that allows similar labels to be assigned to pixels with similar properties in order to solve the problem of inaccurate details during mechanical assembly segmentation owing to the invariant properties of deep learning networks, thus improving the segmentation performance. The SKC-UNet + DenseCRF method was evaluated on three types of datasets containing mechanical assembly segmentation depth images. The results showed that the mean intersection over union (MIoU) of this method reached the optimum value on all the three datasets compared to other semantic segmentation networks. In summary, the proposed network is suitable for mechanical assembly segmentation tasks and can be applied to product assembly monitoring.