Editor's Note: Despite the hardware focus system, large image datasets acquired with automated microscopes often have low-quality, out-of-focus images. It is important to use high-precision automated image analysis to obtain a high-quality, unbiased data set. Google researchers have developed deep neural networks to address this need, helping scientists to capture high-quality images. The following is the compilation of the original text.
Many scientific image applications, especially microscopes, produce several T data per day. These applications benefit from recent computer vision and deep learning. As we work with biologists on robotic microscopy applications, we have learned that combining high-quality images of noise removal from signals into a data set is a daunting and important task. In addition, we realize that many scientists may not write code, but they still like to use deep learning when analyzing scientific images. One of the special problems we can solve is dealing with out-of-focus images. Even with the most advanced autofocus systems on the microscope, improper configuration or incompatible hardware can cause image quality problems. Automated evaluation of focus quality enables image detection, troubleshooting and deletion.
Secured from deep learning
In Assessing Microscope Image Quality with Deep Learning article, we trained a neural network depth, focus on quality microscopy imaging to evaluate the results better than the previous method. We also integrate the pre-trained TensorFlow model with plug-ins in Fiji (ImageJ) and CellProfiler, two advanced open source scientific image analysis tools that can be used through a graphical user interface or by calling scripts.
The basics of the machine learning project workflow are described in published articles and open source code (TensorFlow, Fiji, and CellProfiler): assembling a training data set (we defocus 384 cell focus images to avoid data sets that need to be manually labeled) ), using a data-enhanced training model to evaluate generalization (in our case, an invisible cell type is obtained through an additional microscope) while deploying a pre-trained model. Previous tools for identifying image focus quality typically required the user to manually examine the image of each data set to determine the critical point between the focus and the out-of-focus image. Our pre-trained models do not require user-set parameters and can be used to more accurately assess focus quality. To enhance interpretability, our model evaluates the focus quality of an 84×84 pixel block, the colored border in the image above.
What about images without target objects?
An interesting challenge we encountered was that many image patches were often "blank", ie without a target object, a situation where there is no concept of focus quality. Instead of explicitly labeling these "blank" patches, we let the model identify them, but instead configure the model to predict the probability distribution of the defocus level so that it learns to express the uncertainty in these blank patches.
What to do next?
Scientific image analysis methods based on deep learning will enhance accuracy, reduce manual manual tuning, and possibly bring new discoveries. It is clear that the sharing and effectiveness of data sets and models, as well as the installation of tools, demonstrate that the widespread implementation of such tools is important in all areas.
RAM/RFM electric heating capacitors
RAM/RFM Electric Heating Capacitors
Electric Heating Capacitor,Film Heating Capacitor,Electric Capacitor Bank,Induction Heating Capacitors
YANGZHOU POSITIONING TECH CO., LTD. , https://www.cnchipmicro.com