WResUNet: an instance segmentation method based on two-way residual networks and asymmetric convolution

At present, many segmentation networks can generate high-quality masks for various types of images, however, their feature expressiveness is limited when handling images with many instances, irregular shapes, or dense arrangements. This limitation restricts further improvement in segmentation accuracy. To address the above problems, we propose an instance segmentation method, WResUNet, based on two-way residual networks and asymmetric convolution, which is based on an encoder-decoder framework. The encoder backbone consists of multiple GRblock modules, which are implemented by fusing maximal pooling and convolution operations in parallel to reduce the semantic loss caused during resolution reduction. In addition, it introduces 2 sets of asymmetric convolutions to form a residual block to perform feature extraction on the fused feature maps and share constant connections. In order to speed up the inference and reduce the number of parameters, the GRblock module also introduces channel attention, spatial attention mechanism, and 1×1 convolutional blocks for further optimization. The decoding part of WResUNet employs a primary decoder DeC and a secondary encoder ADeC. DeC undertakes the higher-order feature maps parsing, while ADeC is used for recovering the model's shallow detail semantic information, which better assists the primary decoder in recovering the more precise semantic information. more precise semantic information. Through experiments on the CVPPP and KOMATSUNA datasets, it has been shown that WResUNet outperforms U-Net, ResUNet, UNet++, and other advanced methods, demonstrating better segmentation performance.