Copyright Protection of Images Based on Large-Scale Image Recognition Koichi Kise, Satoshi Yokota, Akira Shiozaki Osaka Prefecture University
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 2
DEMO Large-Scale Image Recognition
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 4
Image Recognition with Local Descriptors Database Images are represented by local descriptors (feature vectors) Typically a few hundreds to a few thousands are extracted Image 1 Image 2 Image 3 Image 4 Query 5
Database Image 1 Image 2 Image 3 Image 4 Query Image Recognition with Local Descriptors Nearest Neighbor Search no. of votes max. votes 6 Nearest vector in the DB Nearest vector in the DB
Pros and Cons Extraction of descriptors a few seconds / query Matching By brute-force NN search with the DB of 1,000 images 315 seconds / query PROS CONS Robust to occlusion and variations Locality and stability of descriptors Large number of expensive descriptors 7
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 8
Our Research Topics Extraction of Descriptors Matching (with 10,000 images) BaselineSoftware2.3 sec.Brute-force3,000 sec. Existing Technologies ANN (Approximate Nearest Neighbor) 50 ms 60,000 times faster Our MethodsGPU & CUDA0.57 sec.Hash & Cascade0.78 ms 4 times faster60 times faster 9 Application Copyright Protection
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 10
Copyright Protection and Digital Watermarking Digital Watermarking is a technology to embed information (copyright notice) to original digital data. Embedded information (watermark) is often unnoticeable. We focus on digital images as original data. Methods: Non-blind Blind 11 Copyright: A Purchaser: B Copyright: A Purchaser: B original watermarked
Non-Blind and Blind Watermarking Non-blind watermarking Blind watermarking History From non-blind to blind watermarking Non-blind watermarking needs matching of images spoils the scalability Blind watermarking has been the focus of interest 12 watermarked image & originalimage Copyright: A Purchaser: B Copyright: A Purchaser: B watermarked image Copyright: A Purchaser: B Copyright: A Purchaser: B
Limitations of Blind Watermarking StirMark distorts images imperceptibly severely damages the wartermark 13 OriginalStirMark
StirMark: Some Image Transformations 14 corner shift cropping punch / pinch original fraction of pixel displacement
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 15
Proposed Method: Ideas How do we solve the problem of blind watermarking? Cancel the effect of StirMark, especially geometric transformations Normalization could be done by comparing with the original image Blind to non-blind 16 StirMark’ed OriginalStirMark’ed normalization normalized StirMark’ed
Proposed Method: Ideas How to solve the problem of non-blind watermarking? Matching problem solved by the efficient and robust image recognition How about the normalization? Image recognition with local descriptors provides matching of descriptors as a side effect helpful for normalization 17
Proposed Method: Ideas To solve the problem of StirMark Image normalization is required to detect the watermark Normalization could be done by comparing with the original image Non-blind watermarking Efficient and robust image recognition overcome the difficulty of matching in non-blind watermarking precise matching of local descriptors (feature points) is obtained as a side effect of recognition This helps us to normalize the StirMark’ed image to detect the watermark 18
Proposed Method: Processing Steps 19 STEP1: Recognition Image Database query image point matching STEP2: Normalization STEP3: Detection Embedded Watermark
Proposed Method: Processing Steps 20 STEP2: Normalization Embedded Watermark STEP3: Detection Local Descriptors: PCA-SIFT Matching: ANN Local Descriptors: PCA-SIFT Matching: ANN STEP1: Recognition Image Database query image point matching
Outline Demo Image Recognition with Local Descriptors Our Research Topics Copyright Protection and Digital Watermarking Proposed Method Experiments Conclusion 21
Experiments: Overview Experiment 1 Robustness against StirMark with default setting (imperceptible change) Experiment 2 Robustness against perceptible changes Experiment 3 Robustness against watermarking 22
Experiments: Overview Experiment 1 Robustness against StirMark with default setting (imperceptible change) Experiment 2 Robustness against perceptible changes Experiment 3 Robustness against watermarking 23
Experiments: Dataset & Task 24 10,000 images 100 Base set WatermarkStirMark query images Input : query image Task: find its original from 10,000 images Input : query image Task: find its original from 10,000 images
Experiment 1 25 imperceptible 100 random application with default setting 10,000 images 100 StirMark query images 10,000 images 100 query images watermarkingwithoutwith accuracy100%99.49% time (per query)217ms216ms Watermark with StirMark without 10,000
Experiment2 4 image transformations cropping, fraction of pixel displacement, punch/pinch, corner shift perceptible 26 perceptible 4 trans. 43 application in total 10,000 images 100 StirMark query images Watermark 10,000 images 100 StirMark query images with without 4,300
97% Cropping: 100 ; Accuracy: 97%
Cropping: Original Image
97% Cropping: 70; Accuracy 97% 29
95% Cropping: 50; Accuracy: 95% 30
92% Cropping: 30; Accuracy: 92% 31
Cropping 32 With watermarking Without watermarking Accuracy [%] Cropping parameter
Punch / Pinch Original Image 33
99% Punch / Pinch: 1; Accuracy: 99% 34
99% Punch / Pinch: 25; Accuracy: 99% 35
97% Punch / Pinch: 50; Accuracy: 97% 36
50% Punch / Pinch: 100; Accuracy: 50% 37
Punch / Pinch 38 Accuracy [%] Punch / Pinch Parameter With watermarking Without watermarking
Conclusion A non-blind framework for copyright protection of images has been proposed Efficient image recognition technology with local descriptors Fundamental experiments on image recognition Image recognition is robust to StirMark attacks and watermarking Future Work Completion of overall method 39
Copyright Protection of Images Based on Large-Scale Image Recognition Thank you for your attention
Image Recognition: Local Descriptors PCA-SIFT (PCA Scale-Invariant Feature Transform) 36 dimensions Invariant to scale and rotation changes 41
Image Recognition: Matching ANN(Approximate Nearest Neighbor) Tree structure Feature Space Feature Vector
ANN : Feature Vector of a query
ANN Nearest Neighbor Search for
ANN NN Search for Approximate Result This cell is not searched
ANN more approximation with a smaller circle Wrong Correct Errors caused by approximation NN Search for Approximate
NNS accuracy vs. Image recognition rate Image recognition rate[%] NNS accuracy[%] ANN dist E2LSH
Corner Shift Original Image 48
99% Corner Shift: 1; Accuracy: 99% 49
98% Corner Shift: 100; Accuracy: 98% 50
96% Corner Shift: 150; Accuracy: 96% 51
74% Corner Shift: 225; Accuracy: 74% 52
Corner Shift 53 Accuracy [%] Corner Shift Parameter With watermarking Without watermarking
Fraction of Pixel Displacement Original Image 54
99% Fraction of Pixel Displacement: 1; Accuracy 99% 55
94% Fraction of Pixel Displacement: 3 Accuracy: 94% 56
81% Fraction of Pixel Displacement: 5; Accuracy: 81% 57
58% Fraction of Pixel Displacement: 8 Accuracy: 58% 58
24% Fraction of Pixel Displacement: 10 Accuracy: 24% 59
Fraction of Pixel Displacement 60 Accuracy [%] Fraction Parameter With watermarking Without watermarking