Image grayscale is to downscale a 3-dimensional color image into a 1-dimensional grayscale image. Due to the act of downscaling, the information of a 3-dimensional matrix is represented by a 1-dimensional matrix, and there will inevitably be information loss, which makes it very important to maintain the detail contrast information of the original color image to the maximum extent. In this regard, this paper proposes a two-way contrast retention model and algorithm based on cosine similarity. This model consists of two local contrast retention sub-models based on cosine similarity, and the two sub-models achieve complementary functions, which can play a contrast retention dephasing effect on the regions with large contrast and regions with small contrast in the original color image respectively, so that the total model can play a two-way contrast retention effect. In addition, the model is solved using a parameter discrete search strategy to improve the real-time performance of the algorithm.
Accurate rectal staging has great significance for surgery. However, it is difficult for radiologists to diagnose because of the heavy workload and the shortage of experts. In this paper, we proposed a deep neural network to automatically identify the rectal cancer T-stage. The network combines 3D DenseNet with 3D attention proposal network, which can use the 3D image as training data and take full advantage of the latent relation of each slice. The APN can distinguish slight differences of each rectal cancer T-stage, which use the last convolution layer to create the attention region that can generate multi-scale images. Moreover, the APN can focus on the tumor region and use a rectangle box to identify the related region, which is approximately regarded as network interpretation. The experiment is conducted on MRI images of 254 patients with rectal cancer of T1, T2, T3 stage. In contrast experiment, our proposed method outperforms other methods, so it is indicated that our method is effective.
In this paper, we study the architectures of space-air-ground integration network (SAGIN) proposed by domestic scientific research institutes, and put forward an collaborative federal learning architecture suitable for SAGIN to solve the problems of insecurity and low timeliness caused by traffic backhaul. An anomaly traffic detection method is proposed based on the requirements and characteristics of SAGIN. The problem that it is difficult to manually label and extract features in the traffic of SAGIN is solved through the improvement of deep learning algorithm. The challenge of lack of professionals labeling training set is solved by studying the method of semi supervision. The problem of artificial feature engineering is solved by studying the end-to-end anomaly traffic detection algorithm. Finally, we design a simulation environment for the anomaly traffic detection in SAGIN, and verify the feasibility and advanced nature of the proposed methods.
With the continuous improvement of deep learning algorithms, it has achieved considerable results in the fields of machine vision and natural language processing. In the face of problems that cannot be solved by traditional methods or are not effective, deep learning can achieve more desirable results through its powerful feature learning and mastering the laws of the problem. The lane recognition based on a convolutional neural network is proposed to address the problems of lane crushing of self-driving intelligent vehicles and the accuracy of traffic marker recognition, respectively. The traffic marker recognition model based on YOLO5, by collecting data and completing model training, finally achieves the average lane crush of the self-driving car less than 1 time, and the traffic marker recognition rate reaches 98.5
When the software in embedded system runs, the current may change dynamically in wide ranges. It is difficult to accurately measure the energy consumption of software in embedded system. Some existing methods can provide the functions that measuring the currents of different regions by using different ranges through adjusting shunt resistor or amplification. However, these methods are hard to give the value of current accurately during range switching, and bring some measuring errors. To address this problem, this paper designs hardware and software schemes to measure the current of software in embedded system by using small, medium and large ranges simultaneously. Further, a two-stage calibration method based on machine learning is proposed. And a Measurement System for Energy consumption of software in Embedded system (MSee) is presented. The experimental results show that both the average relative errors of MSee in small, medium and large ranges and the average relative errors of MSee in transitional neighborhood of ranges are better than those existing methods of range switching.
