REAL TIME LOGO RECOGNITION USING YOLO ON ANDROID

Аннотация

Humans can easily detect and identify objects present in an image. The human visual system is fast and accurate and can perform complex tasks like identifying multiple objects and detect obstacles with little conscious thought. For a long time, humans have been trying to make computers understand what is on the images. With the availability of large amounts of data, faster Graphics Processing Unit (GPU)s, and better algorithms, we can now easily train computers to detect and classify multiple objects within an image with high accuracy. The goal of this paper is to implement an object detection model suitable in terms of size and speed to run on an Android device and detect logos in real-time. The proposed approach is based on YOLOv2 (You Only Look Once) state-of-the-art, real-time object detection for logos and this project used the FlickrLogos-32 dataset. The experimental results show that we obtained a final accuracy of 82.3% and a speed of 35 fps (frames per second) on the NVidia GeForce GTX 1070. logos are shown in Figure 1. Logo detection is a challenging object recognition and classification problem as there is no clear definition of what 1. INTRODUCTION constitutes a logo. A logo can be thought of as an artistic expression of a brand, it can be either a A logo is a graphical mark used to identify a (stylized) letter or text, a graphical figure or any company, organization, product or brand. Logos are combination of these. Furthermore, some logos used to represent a company’s name, a particular product or service and are used heavily vin the marketing of products and services. Logos have become an integral part of a company’s videntity and a well-recognized logo can increase a company’s goodwill. A logo usually has a recognizable and distinctive graphic design, stylized name or unique symbol for identifying an organization. It is affixed, included, or printed on all advertising, buildings, communications, literature, products, stationery, vehicles, etc. Logo can be seen anywhere in the surrounding in our vdaily life, such as in the streets, Figure 1: Some figures illustrate that logos supermarkets, on the products or services, on appear everywhere in our surrounding. administrative documents, etc. Examples of different have a fixed set of colors with known fonts while interest, typically defined with bounding boxes. A others vary a lot in color and specialized unknown commonly used region proposal method is Selective fonts. Additionally, due to the nature of a logo (as Search. This algorithm proposes regions of interest by brand identity), there is no guarantee about its context using similarity measures based on color and visual or placement in an image, in reality logos could features. R-CNN method crops and resize each region appear on any product, background or advertising surface. Also, this problem has large intra-class of interest and classifies them using a CNN. The variations e.g. for a specific brand, there exist various original architecture uses a CNN with five logos types (old and new Adidas logos, small and big convolutional layers and two fully connected layers, versions of Nike) and inter-class variations e.g. there although any CNN classifier can be used.Some more exists logos which belong to different brands but look complex methods for object detection include Fast R- similar (see Figure 2). CNN and Faster R-CNN. Fast R-CNN is a method based on R-CNN in which the full image is processed by the convolutional layers and then, regions of the output of the last convolutional layer are cropped and classified. The network is formed by a set of Figure 2: Logo variations exemplar images convolutional layers, fully-connected layers, an Left variations of brands Adidas. Notice, different external region proposal method (typically Selective graphical figures. Right variations of brands Chanel - Search) and a Region of Interest (RoI) pooling layer. Gucci, Vodafone, Target, beats, Bebo and Pinterest. The RoI pooling layer applies max-pooling to each Notice, similar looking logos but belong to different region of interest using a grid of a fixed size (typically brands. 7 × 7). Fast R-CNN also introduces a bounding box 2. RELATED WORK regressor, a layer that outputs a fine-tuned location of bounding boxes. Faster R-CNN is based on Fast R- The problem of logo recognition itself has a CNN but substitutes the external region proposal rich research history. In the 1990’s the problem was methods by a Region Proposal Network (RPN). RPN mainly explored in information retrieval use-cases. is a neural network that generates regions of interest An image descriptor was generated using affine using the features of the output of the last transformations and stored in a database for retrieval. convolutional layer. RPN is formed by a 3 × 3 sliding There were also some neural network-based window that outputs a set of bounding boxes approaches but the networks were not as deep nor the (typically 9) with different