Multi-scale sliding window similar picture detection method

Abstract

The invention discloses a multi-scale sliding window similar picture detection method which is characterized in that interested domains are constructed by using sliding windows of different scales, saliency objects are screened by using a constructed multi-layer convolutional network, more intensive sampling is carried out near the saliency objects, and part of pure-color invalid interested domains are removed; interested domain features are extracted by using the constructed self-adaptive pooling convolution auto-encoder, are compared with stored picture features, and whether candidate pictures are similar to pictures in a stored picture library or not is judged. Objects can be selected more accurately and completely, feature extraction is more accurate, expression is richer, the method can adapt to image stealing behaviors of different scales such as partial screen-shots, the effect of detecting similar images in an omnibearing mode is achieved, behaviors such as infringement are greatly reduced, and the workload of auditors is reduced.

Description

technical field [0001] The invention relates to the field of automatic detection of similar pictures, in particular to a multi-scale sliding window similar picture detection method. Background technique [0002] At present, sellers on the e-commerce website platform mainly use pictures and text to display their products, among which pictures are more intuitive and comprehensive, which are well recognized by buyers. Therefore, picture display has become one of the main display methods for product display. [0003] Due to the huge amount of pictures uploaded on the e-commerce website platform every day, some merchants have stolen other people's pictures. This infringement has greatly damaged the rights and interests of the authors of the original pictures. At the same time, some pirated pictures use screenshots, cutouts, etc., and the original detection method for completely similar pictures cannot be applied. [0004] Patent CN201911037783, a method for recognizing small an...

AI Technical Summary

Problems solved by technology

[0003] Due to the huge amount of daily uploads of pictures on the e-commerce website platform, some merchants have stolen other people's pictures. This infringement has greatly damaged the rights and interests of the original picture author.

At the same time, some pirated images use screenshots, cutouts, etc., and the original detection method for completely similar images cannot be applied

[0004] Patent CN201911037783, a method for recognizing small and medium-sized objects in large-size images based on sliding windows, which divides large-size images into fixed-size images through sliding windows, meets the input requirements of the target detection neural network, and solves the problem that small-size objects cannot be detected due to scaling of large-size images. The problem of detection, but for the specific objects in the picture, it cannot be densely sampled to obtain as much information as possible. At the same time, the ratio of the size of the sliding window is relatively fixed

[0005] Patent CN201810291398, an image multi-scale feature extraction method, through edge detection and corner point detection to obtain the salient area, through more intensive sampling of the salient area to extract more image features, but there are backgrounds, texts, etc. on the image For non-main objects, the area where the salient objects are located cannot be confirmed through edge and corner detection. At the same time, the solid-color background area generated by the sliding window is not filtered, which is easy to cause feature redundancy, and the extracted image gradient direction histogram feature is based on The edge features of the directional gradient can only describe the relevant shape, and cannot describe the overlapping part, and cannot describe other image features such as color

Images