6 results about "Random text" patented technology

The invention provides a text video retrieval method of a fine-grained relation learning network based on energy perception. The method comprises the following steps: giving a query text and a video clip; inputting the query text into a text encoder of the CLIP, inputting the video clip into a text encoder image encoder of the CLIP, and extracting to obtain text embedding and frame embedding; inputting the text embedding and the frame embedding into a fine-grained relation learning network for text enhancement operation to obtain enhanced text embedding; taking the enhanced text embedding as a frame fusion condition to carry out frame fusion operation on the frame embedding to obtain video embedding; calculating the similarity of a text-video pair formed by the query text and the video embedding based on a cosine similarity function; and selecting the text-video pair with the highest similarity as a retrieval output result of text video retrieval. According to the method, the problem of randomness of the random text of single sampling is solved, so that semantic information of text coding is better expanded, and the final retrieval effect is improved.

The invention provides an efficient defense method for poison attack of a pre-training model in a multi-modal scene, and relates to the technical field of defense of poison attack of a deep learning model. The method relates to three stages: in the first stage, before model training, poisoning data in a training set is cleaned and filtered; in the second stage, in model training, a random text pool is used for further detoxification; in the third stage, after model training, clean data are introduced for model fine adjustment; in the first stage of the method, 50% of poisoning samples in a training data set can be filtered out; in the second stage, the efficient random text pool is used for detoxication in the pre-training process, poison samples in a training set can be further removed, and the overall training efficiency is improved; and in the third stage, the influence of a random text pool on the accuracy of the training model can be further made up through post-fine-tuning detoxification, and meanwhile, external clean data is introduced, so that the detoxification effect is further improved.

The present invention discloses an OCR training sample generation method, device and system, which relates to the field of computer vision. The method includes: a text contour extraction step, where all text contours are extracted based on the original image, an erasure area mask is determined by combining the erasure area coordinates, and a repaired area mask is obtained; an image repair and filling step, where image repair and filling are performed according to the repaired area mask and the pixel information around the repaired area to obtain a background template after the erased text; a random text generation step, where random text is generated in each generation area, thereby obtaining a new sample picture and a corresponding annotation information file. The present invention combines technologies such as text contour extraction algorithms and image repair, makes full use of the background information of the original picture, generates high-quality training pictures, and at the same time generates an annotation file corresponding to the picture (including text content and position information), eliminating the cumbersome and laborious annotation work and can be directly used for OCR model training.

The embodiment of the application provides an image generation method, a model training method and a text bolding method, and belongs to the technical field of image processing. The image generation method comprises the following steps: randomly selecting one from a plurality of preset image sizes as the size of a first image, and generating the first image containing random text; performing bolding on the random text in the first image based on preset bolding parameters to obtain a second image; wherein the preset bolding parameters comprise a preset bolding direction and a preset bolding value; and zooming the first image and the second image to a target image size. The pre-set deep learning network model is trained by using the processed first image and second image, which can improve the generalization ability of the model and ensure the readability and aesthetic degree of the text in the screen content image output by the model.

This invention discloses a zero-shot adversarial robust model, method, and computer device based on text information enhancement. First, text descriptions and random text are generated to enhance semantic expression diversity. Features of images and different texts are extracted using a visual language model. Adversarial examples are iteratively generated using image features and text description features, and used for adversarial training. Cross-entropy loss is calculated using the image features and text description features of the adversarial examples to train the adversarial model. The cosine similarity between random text features, adversarial example features, and clean example features is calculated separately, and the two are aligned using a loss. To maintain zero-shot performance on clean examples, the cosine similarity between random text features and clean example features in the original and target models is calculated, and the two are aligned using a loss. The three losses are integrated to obtain a zero-shot adversarial robust model based on text information enhancement. This method can improve the adversarial robustness and generalization performance of the model under zero-shot conditions.

The invention relates to a speech synthesis technology, and discloses a speech synthesis method based on a generative adversarial network, which comprises the following steps: carrying out feature analysis on text input to obtain audio features; inputting the random text into a generator in a pre-constructed generative adversarial network to obtain an initial voice signal; using a discriminator in the generative adversarial network to carry out loss evaluation on the audio features and the initial voice features; if the loss evaluation does not meet a preset condition, optimizing the generator and the discriminator according to a loss evaluation result to obtain a target generative adversarial network; calculating the random text by using a generator in the target generative adversarial network to obtain a target voice signal; and performing voice reading on the target voice signal to obtain a playable voice. The invention further provides a speech synthesis device based on the generative adversarial network, electronic equipment and a storage medium. The method can improve the quality of generated voice and the synthesis efficiency.