高效和准确的场景文本(efficient and accuracy scene text,EAST)检测算法速度快且结构简单,但是由于文本结构的特殊性,导致在检测中尺寸较小的文本会被遗漏,而较长的文本则完整性较差。针对EAST算法存在的问题提出一种新的自然场景文本检测模型。该方法利用自动架构搜索的特征金字塔网络(neural architecture search feature pyramid network,NAS-FPN)设计搜索空间,覆盖所有可能的跨尺度连接提取自然场景图像特征。针对输出层进行修改,一方面通过广义交并比(generalized intersection over union,GIOU)作为指标提升边界框的回归效果;另一方面通过对损失函数进行修改解决类别失衡问题。输出场景图像中任意方向的文本区域检测框。该方法在ICDAR2013和ICDAR2015数据集上都取得了较好的检测结果,与其他文本检测方法相比,检测效果也得到了明显提升。
This paper proposes a novel high-power supply rejection ratio（high-PSRR） high-order curvature-compensated CMOS bandgap voltage reference（BGR） in SMIC 0.18 μm CMOS process. Three kinds of current are added to a conventional BGR in order to improve the temperature drift within wider temperature range, which include a piecewise-curvaturecorrected current in high temperature range, a piecewise-curvature-corrected current in low temperature range and a proportional-to-absolute-temperature T^（1.5） current. The high-PSRR characteristic of the proposed BGR is achieved by adopting the technique of pre-regulator. Simulation results shows that the temperature coefficient of the proposed BGR with pre-regulator is 8.42x10^（-6）′ /℃ from - 55 ℃ to 125 ℃ with a 1.8 V power supply voltage. The proposed BGR with pre-regulator achieves PSRR of - 123.51 dB, - 123.52 dB, - 88.5 dB and - 50.23 dB at 1 Hz, 100 Hz, 100 kHz and 1 MHz respectively.
