基于轻量化网络和密集光流法的精子活力检测

基于便携式家用精子检测仪的研发需求,研究了轻量化卷积神经网络在精子活力检测方面的应用.利用Farneback光流算法提取出不同帧间距的精子视频的密集光流帧图像,并通过多通道图像叠加的方式将其与原始视频帧图像进行叠加.把叠加后的图像作为轻量化卷积神经网络ShuffleNet的输入图像,用于检测视频中精子的活力水平.实验结果表明：使用ShuffleNet能在基本不降低检测精度的前提下显著降低网络整体的计算量和模型所占用的内存,更适用于嵌入式和移动设备.此外,采用多通道叠加密集光流帧和原始帧图像作为输入图像,相较于单一的原始帧图像,有效提升了网络模型的性能.

Sperm motility detection based on lightweight network and dense optical flow method

Based on the research and development needs of portable home sperm detection, the application of lightweight convolutional neural network in sperm motility detection was studied in this paper. The Farneback optical flow algorithm was used to extract the dense optical flow frame images of sperm videos with different frame intervals, which were stacked with the original video frame images by multi-channel image superposition. These stacked images were used as the input of the lightweight convolutional neural network ShuffleNet to detect the motility level of sperm in the video. Experimental results showed that by use of ShuffleNet the overall computation cost of the network and the memory space could be reduced significantly without reducing the detection accuracy, which was more suitable for embedded and mobile devices. Furthermore, compared to utilizing only original frame pictures as input, the performance of the network model was effectively improved with the adoption of using multi-channel superposition of dense optical flow frame images and original frame images as input.