Deep learning and computer vision

Deep learning has a good performance in dealing with more information-rich tasks, which is very suitable for computer vision tasks, and large-scale datasets and the powerful ability of deep learning networks provide a broad development space for computer vision.

Deep learning

Deep learning is a field in machine learning which is the study of multiple data such as images, videos, text, sound, etc. through deep understanding and learning from datasets or sample libraries. Deep learning plays a great role in search technology, machine translation, computer vision, natural language processing, personalized recommendations and many other areas of technology.

Deep learning can be divided into three categories in terms of research content, namely, neural network systems based on convolutional computation, often called convolutional neural networks, and self-coding neural networks based on multi-layer neurons, and deep confidence networks three categories. With the deepening of deep learning research, researchers are gradually combining different methods and different training steps to achieve better training results. Compared with traditional methods, deep learning has more parametric models, so the amount of data involved in the training is larger, and the training of the model is more difficult, but the training will achieve better results.

Deep Learning and Computer Vision

Deep learning has a good performance in dealing with tasks that are richer in information, which is very suitable for computer vision tasks, and large-scale datasets and the powerful ability of deep learning networks provide a broad space for computer vision. With the addition of deep learning, computer vision has gradually spread from the initial image transformation, image coding and compression, image enhancement and restoration, image segmentation, image description and so on to more complex areas. The most common image classification and recognition in life, such as face recognition, fingerprint recognition, license plate recognition and so on.

Localized Convolutional Neural Network

Localized Convolutional Neural Network (Region-CNN, R-CNN), is the first algorithm to apply deep learning to target detection, and the accuracy of R-CNN’s target detection has been dramatically improved compared to its previous algorithms.

In traditional target detection, we first circle all the region frames that may be the target objects on the image, then extract features from these region frames and classify them by image recognition methods, and then output the classified regions by non-extremely large value suppression methods.R-CNN preserves the idea of traditional target detection, and retains the methods of using region frames to extract features, classify the images, and suppress non-extremely large values. But the difference is that the traditional feature extraction method is replaced with deep convolutional network feature extraction method. The specific steps of R-CNN are shown below.

Common datasets

Data sets are an indispensable part of deep learning, deep learning is all based on the information carried by a large amount of data within the dataset, the larger the amount of dataset used for training, the better the training results obtained may be. The datasets required for computer vision are relatively large and complicated to collect personally, so there are many publicly available datasets on the web that researchers can use for learning.

Your task

What are some aspects of deep learning in computer vision? This can be illustrated with an example.

Share your thoughts and ideas in the comments below.