|
[1910.11107] Streaming Networks: Enable A Robust Classification of Noise-Corrupted Images
To summarize, we emphasize that the main achievement of our work is the introduction of a simple and uncostly method to increase conv net robustness against the noise without using complex data generation techniques or sophisticated learning algorithms
Abstract: The convolution neural nets (conv nets) have achieved a state-of-the-art
performance in many applications of image and video processing. The most recent
studies illustrate that the conv nets are fragile in terms of recognition
accuracy to various image distortions such as noise, scaling, rotation, etc. In
this study we focus on the problem of robust recognition accuracy of random
noise distorted images. A common solution to this problem is either to add a
lot of noisy images into a training dataset, which can be very costly, or use
sophisticated loss function and denoising techniques. We introduce a novel conv
net architecture with multiple streams. Each stream is taking a certain
intensity slice of the original image as an input, and stream parameters are
trained independently. We call this novel network a "Streaming Net". Our
results indicate that Streaming Net outperforms 1-stream conv net (employed as
a single stream) and 1-stream wide conv net (employs the same number of filters
as Streaming Net) in recognition accuracy of noise-corrupted images, while
producing the same or higher recognition accuracy of no noise images in almost
all of the tests. Thus, we introduce a new simple method to increase robustness
of recognition of noisy images without using data generation or sophisticated
training techniques.
‹ Figure 1. Intensity Slices (The conv nets and the primate brain) Figure 2. One stream simple comv net (The conv nets and the primate brain) Figure 3. Streaming Network architecture (The conv nets and the primate brain) Figure 4. Adding random zero noise to Eurosat original data. (The conv nets and the primate brain) Figure 5. Tests of Cifar10 dataset using Adam optimizer with 0.0005 learning rate for 1-stream conv net and 5-stream Streaming Net. Green line illustrates average prediction accuracy of no noise data by 5-stream Streaming Net, red line illustrates average prediction accuracy of no noise data by 1-stream network. Blue dotted line illustrates noise-corrupted data prediction accuracy (one sample run) by 5-stream Streaming Net and blue solid line is 7-point moving average smoothing. Orange dotted line illustrates noisecorrupted data prediction accuracy (one sample run) by 1-stream network and orange solid line is 7-point moving average smoothing. (The conv nets and the primate brain) Figure 6. Tests of Cifar10 dataset using Adam optimizer with 0.0001 learning rate for 1-stream conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ?? (The conv nets and the primate brain) Figure 7. Tests of Cifar10 dataset using Adam optimizer with 0.0005 learning rate for 1-stream wide conv net and 5-stream Streaming Net. The green, blue and blue dotted lines have the same meaning as in Fig. ??, while other lines, corresponding to the same shape and color lines in Fig. ??, refer to 1-stream wide conv net’s performance. (The conv nets and the primate brain) Figure 8. Tests of Cifar10 dataset using Adam optimizer with 0.0005 learning rate for 1-stream wide conv net and 5-stream Streaming Net. The green, blue and blue dotted lines have the same meaning as in Fig. ??, while other lines, corresponding to the same shape and color lines in Fig. ??, refer to 1-stream wide conv net’s performance. (The conv nets and the primate brain) Figure 9. Tests of Eurosat dataset using Adam optimizer with 0.0005 learning rate for 1-stream conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ?? (The conv nets and the primate brain) Figure 10. Tests of Eurosat dataset using Adam optimizer with 0.0001 learning rate for 1-stream conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ??. (The conv nets and the primate brain) Figure 11. Tests of Eurosat dataset using Adam optimizer with 0.0005 learning rate for 1-stream wide conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ??, while other lines, corresponding to the same shape and color lines in Fig. ??, refer to 1-stream wide conv net’s performance. (The conv nets and the primate brain) Figure 12. Tests of Eurosat dataset using Adam optimizer with 0.0001 learning rate for 1-stream wide conv net and 5-stream Streaming Net. The green, blue and blue dotted lines have the same meaning as in Fig. ??, while other lines, corresponding to the same shape and color lines in Fig. ??, refer to 1-stream wide conv net’s performance. (The conv nets and the primate brain) Figure 13. Tests of Eurosat dataset using Adam optimizer with 0.0001 learning rate for 1-stream wide conv net and 10-stream Streaming Net. The green, blue and blue dotted lines have the same meaning as in Fig. ??, while other lines, corresponding to the same shape and color lines in Fig. ??, refer to 1-stream wide conv net’s performance. (The conv nets and the primate brain) Figure 14. Tests of UCmerced dataset using Adam optimizer with 0.0001 learning rate for 1-stream conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ??. (The conv nets and the primate brain) Figure 15. Tests of UCmerced dataset using Adam optimizer with 0.0001 learning rate for 1-stream wide conv net and 5-stream Streaming Net. The lines have the same meaning as in Fig. ??. (The conv nets and the primate brain) Figure 16. Average recognition accuracy curves (learning curves) for the Eurosat dataset with a learing rate 0.0001. (Streaming Net with more streams)›
|
|