<p>Microlens recognition in industry automation is currently a challenge due to complex background, dim target, and blurred edge. In this study, a multiscale feature hierarchical convolutional attention network (MHC-Net) is proposed to realize the accurate recognition of microlens for the automatic optical detection. The MHC-Net is designed with encoder-decoder structure, which includes down-sampling, up-sampling, multiscale feature fusion block (MFB), spatial-attention neighborhood enhancement module (SNEM), depth-separable atrous asymmetric convolution module (DAAM). The MFB adaptively selects and fine fuses of both high-dimensional and low-dimensional features, which enhances the prominence of microlens. The SNEM fuses neighborhood information and spatial attention for enhancing the feature representation of semantic information. The DAAM integrates depth-separable convolution, atrous convolution, and asymmetric convolution, which can effectively extract multi-scale features to capture location information of small microlens. The MHC-Net achieves 94.94% in F1 score and 90.36% in Mean Intersection over Union on a self-built high-quality pixel-labeled microlens dataset, and behaves better performance on public DAGM 2007 dataset. Compared with state-of-the-art segmentation networks, our model achieves better segmentation results, while significantly reducing the computational complexity.</p>