Summary of video frames to retrieve text with Depthwise Separable Convolutions

Text in natural images and video is an important source of information that is used for many real-world applications. Nowadays, text recognition in video and images is a popular and well-known issue that attracted a lot of attention. In applications such as Video browsing, event detection, text retrieval, text recognition and extracting. There are several methods for recognizing text that instead of the entire input frame, which is both text and non-text, they only use frames containing text, so summarizing video frames based on the text is an initial process for detecting and extracting text that needs to accurately and quickly detect non-text frames of the text, which reduces the computational cost and also avoids the high cost of text recognition methods that give non-text video frames as inputs. The text and non-text classification algorithm should be implemented in a short time and with high accuracy and recall. The proposed method has been examined and tested on the TextDis standard dataset that has been specifically designed to summarize the collection of images based on non-text and text. In addition, the main objective is to summarize video frames containing persian texts. Therefore, two sets of data sets have been specially designed to provide the first set of natural images collected from the internet and the second set of video frames that are for the first time the total has been developed.