Search

The two-phase flow in a microchannel consists of liquid–liquid and gas–liquid material components. The automatic recognition of flow patterns using deep learning approaches has been emerging. This study aimed to improve the recognition accuracy of flow patterns in the two-phase flow images. The different convolutional kernels in the GoogLeNet algorithm extracted the image features with different scales. In order to strengthen the important channel and spatial features, this paper proposes the combined five-layer Coord attention and GoogLeNet algorithm to enhance the accuracy of the new algorithm. The optimized algorithm model was derived from image datasets with different liquid–liquid two-phase flows (NaAlg–Oil, GaInSn–Water), and its accuracy was 95.09% in training and 98.12% in testing. This new model was also applied to predict the flow patterns, with a recognition accuracy of more than 97% in both the liquid–liquid and gas–liquid two-phase flows (water–soybean oil, water–lubricating oil, and argon–water).

This paper presents the optimization design of a miniaturized five-element wide-angle fisheye lens using a deep learning algorithm. Zemax optical design software was used to simulate and optimize the wide-angle fisheye lens. A deep learning algorithm helped to find the best combination of different lens materials. We first used six lens elements as an initial configuration to design miniaturized wide-angle fisheye lenses using the optimization process. The optical system components were gradually decreased to five lens elements. Both OKP4HT and polymethyl methacrylate (PMMA) plastic aspheric lenses were selected to replace the second spherical glass lens in the original design. We propose two types of wide-angle fisheye lens designs with four spherical lenses and one aspheric lens. The results for these designs indicated a viewing angle of 174°, a total length of less than 15 mm, a spot size of less than 6 μm, lateral color within ±1 μm, field curvature within ±0.02 mm, and F-θ distortion of ±3.5%. In addition, the MTF value was larger than 0.4 at the spatial frequency of 100 cycles/mm.