Shulou(Shulou.com)06/02 Report--

At present, computer vision has become an interdisciplinary field. Computer vision originated from neural network technology around 1980, but it has not been truly commercialized on a large scale until recently. Large-scale funding has prompted more companies focused on computer vision to emerge, and these companies are constantly upgrading their original business models through computer vision technology in different fields. As a technology inspired by the human visual cortex, are we now at a stage where machines can detect or classify objects as well as or better than human vision?

Vision Technology: AI Nose Print Recognition

Recently, Vision Technology launched AI nose print recognition solution, which was first applied to dog identity authentication. Dedicated to pet identification. Different from other dog identification methods such as pupil and face, the nose pattern is selected as the key feature of recognition. Similar to human fingerprints, dog nose prints are unique and stable, that is, there are no two dogs with identical nose prints, and the nose prints of the same dog do not change with age. The owner only needs to aim at the nose of the dog for simple capture or video recording. The system detects the nose of the dog, locates the key points of the nose print, imports the extracted nose print depth map information into the background database, and can generate a dedicated ID card for the dog. At present, the nose print recognition technology can achieve 1:1 comparison of dogs, and the accuracy rate in pilot scenes is 95% when the error rate is 1 in 10,000;

Polar Chain Technology: Video Recognition

At present, there are still many difficulties and challenges in video Face Recognition, such as poor video image quality and small light problem of face image. Polar Chain Technology proposes to recognize faces in scenes with four modules.

1. Video structuring, dividing video into shots. Generally, the method of combining global features with local features is adopted. Global features detect the distribution mutation of global color, and then use the tracking results of Face Recognition obtained by local features and the discontinuity of tracking trajectory to judge whether the video has shot switching. Tracking to determine shot cuts has a great advantage because similar algorithms are used in subsequent steps, so the algorithm required for this step can be reused.

2. Face track extraction. After the shot segmentation is completed, face trajectory extraction can be performed in a single shot. Accuracy and speed should also be considered in the trajectory extraction algorithm. To achieve the balance between speed and accuracy, there are two ways: one is interval sampling or frame-by-frame processing, the other is detection & tracking cooperation.

3. Face Recognition. With the face trajectory, you can start face recognition. However, before inputting face data into the deep network, it needs to be transformed and processed. One part of the transformation is very important for the face, especially in consumer video, and that is the alignment of faces. Face alignment is a process of restoring and correcting face images with various poses to front faces by using feature points detection and location. In the algorithm framework, we need to add face quality evaluation algorithm to filter low-quality face images and ensure the accuracy of face data. On the premise of sufficient samples, the trained model can be used to extract features from face samples. When testing, after the face is detected in the video, it is input into the generated feature vector, and the feature vector interacting with the face is matched, so as to find the closest sample in the feature space.

4. Fusion of identification results. The above-mentioned Face Recognition is for the single-frame recognized pictures, and the system recognition results mentioned above are for the whole face trajectory. Therefore, finally, the result of Face Recognition needs to be fused with the whole face trajectory to obtain the recognition result of the whole trajectory.

Shangtang Technology: Facial Image Processing

Recently, a team of researchers from Shangtang Technology, the Chinese University of Hong Kong and the University of Hong Kong proposed a new framework called MaskGAN for diverse and interactive facial manipulation. The main idea is that semantic masks act as appropriate intermediate representations for flexible facial manipulations, giving them fidelity. MaskGAN has two main components:

1. dense mapping

2. Editing Behavior Simulation Training

Specifically, the dense mapping network learns a free-form style mapping between the user-modified mask and the target image to achieve different generation results.

Weizmann Institute of Science, Israel: Image Separation

This month, researchers at Israel's Weizmann Institute of Science developed a new technique called Double-DIP, which allows systems to edit images through deep learning without large amounts of training data, separating what people want from what they don't want in images. The study is based on a hybrid image restoration technique called Deep Image Prior (DIP), so the researchers call their new method of separating images Double-DIP. The results of DIP technology research were submitted on July 18, 2018 in arxiv, entitled "Learning Hybrid Sparsity Prior for Image Restoration: Where Deep Learning Meets Sparse Coding."

University of Lübeck: A New Approach to Medical Image Generation

At present, GAN application in medical research still faces a major challenge. Deep learning algorithms need to be trained on high-resolution images to produce the best predictions, but synthesizing such high-resolution images, especially 3D images, requires a lot of computational power. Researchers from the Institute of Medical Informatics at the University of Lübeck have come up with a new approach that can significantly reduce hardware configuration requirements. The researchers broke down the process of image generation into several stages: first generating a low-resolution image with GAN, and then generating a small number of detailed images at the correct resolution. Through experiments, the researchers found that this method not only generated realistic high-resolution 2D and 3D images, but also maintained the same expense regardless of image size.

Summary:

Before the advent of deep learning technology, many applications encountered bottlenecks and progress was slow, with only approximate accuracy improvements every year. But with the advancement of deep learning, the development of computer vision has undergone a huge leap, and the continuous upgrading of technology has also spawned a series of cross-industry applications. With the entry of mainstream technology giants, the field of computer vision has been bustling, but if you want to create some new applications and award-winning application capabilities, I am afraid there is still a long way to go.

https://www.toutiao.com/a6721532654291780108/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.