We are also developing highly potential products, including:
The pupil size or changing speed can indicate both one’s physical condition and mental state. If the pupil contraction or pupil reaction rate is abnormal, it may mean some symptoms such as glaucoma, diabetic retinopathy, general anesthesia, or Horner syndrome. On the other hand, the dilation of the pupil may be stimulated by mental conditions, for example, excitement and pleasure, whereas the contraction of the pupil may be caused by stress, discomfort, and so on. This paper proposes an algorithm that can calculate the pupil size based on Convolution Neural Network (CNN). Usually, the shape of the pupil is not round, and 50% of pupils can be calculated using ellipses for the best fitting. This paper uses the major and minor axis of an ellipse to represent the size of pupils and use the two parameters as the output of network. Regarding the input of network, the data set is from videos (continuous frames). If taking each frame from the videos and using these to train the CNN model, it may cause overfitting because the images are too similar. To avoid this problem, this paper uses data augmentation and calculates the structural similarity to ensure that the images have a certain degree of difference. For optimizing the structure of network, this paper compares the mean error with changing the depth of the network and the field of view (FOV) of the convolution filter. The result shows that both deepening the network and widening the FOV of the convolution filter can reduce the mean error. According to the results, the mean error of the pupil length is 7.63% and the pupil area is 14.68%. It can operate in low-cost mobile embedded systems at 36 frames per second, so the low-cost systems can be used for pupil size prediction.
We dedicated on developing auto bird-detection system based on artificial intelligence. The detection system additionally equips laser for driving away wild birds to stop transmission of virus and to prevent avian influenza. Our target is small-object detection and we utilized data augmentation, curriculum learning and hard negative mining methods and try to improve the performance particularly for small birds. In our study, we verified in the same environment by using several strategies. After collecting massive bird data, we carried out an AI-laser system and performed in real environment.
This research is based on the generative model which using deep learning framework. The applications of this technique can be involved in furniture, tile, wallpaper, Interior decoration, and other industries. The proposed method uses specific texture samples to automatically generate textures that are visually similar to the original samples, but with variation and randomness. The system structure is shown in Figure 1 (a), which includes a VAE (Variational autoencoder) and a GAN (Generative adversarial nets). The task of VAE is generating rough textures, on the other way, GAN is adding high frequency detail on the rough texture which generated by VAE. We constructed the VAE by using residual connections and batch normalization layers to accelerate training and improve image quality, and the design of GAN is based on WGAN-GP.
The training flow is shown in Figure 1 (a). Two types of image are generated by VAE, one is the “reconstructed image” (marked with a red frame in the figure), obtained by passing an original sample through the decoder and encoder; another is “generated image” (marked with a yellow frame in the figure), which is obtained by directly using decoder to decompress a random tensor sampled from Gaussian distribution. For the both types of image, we input they into the G (Generative network) to get the detailed results, and then input the results into the D (Discriminative network) to get the scores of these images and then use them to calculate the “adversarial loss”. In addition, the detailed result and original sample of reconstructed image are input into the trained VGG19 (A classical CNNs model) to get their feature maps, and then calculate the “reconstruction loss” by using L2 distance of these two feature maps. The process of inference stage is shown in Figure 1 (b). A detailed generated texture can be obtained only by passing a random tensor through the decoder and G.
Figure 2 shows the training samples (top) and the randomly generated images (bottom). It shows the performance of texture generation of wood and stone. This system can learn the multi-scale characteristics from texture samples and use them to generate highly simulated textures. Although the results cannot be as detailed as the original samples, this system still has a certain level and potential.
The Inverse Tone-Mapping Operator (ITMO) can extend the dynamic range of standard dynamic range (SDR) images to high dynamic range (HDR), helping the majority of SDR images to be converted into HDR images, thereby exerting the performance of HDR displays and improving image quality. However, several existing ITMOs are designed based on different concepts or purposes, and different image content may be suitable for different ITMOs.
Therefore, this study attempted to create an optimal ITMO prediction system based on supervised learning, the system was trained by artificial intelligence (AI) algorithms which are good at image content analysis. In this research, 8 types of ITMO were evaluated psychophysically through a best image labeling experiment and a categorical judgment experiment. 6 types of machine learning methods (SVM, MLP, KNN, etc.) and 5 types of CNN network architectures (VGG16, ResNet-50, etc.) were tested for the prediction system. The results show that the AlexNet model trained on grayscale images can accurately predict the optimal ITMO.
In this study, we present a new way to predict the Zernike coefficients of optical system. We predict the Zernike coefficients through the function of image recognition in the neural network. It can reduce the mathematical operations commonly used in the interferometers and improve the measurement accuracy. We use the phase difference and the interference fringe as the input of the neural network to predict the coefficients respectively and compare the effects of the two models. In this study, python and optical simulation software are used to confirm the overall effect. As a result, all the Root-Mean-Square-Error (RMSE) are less than 0.09, which means that the interference fringes or the phase difference can be directly converted into coefficients. Not only can the calculation steps be reduced, but the overall efficiency can be improved and the calculation time reduced. For example, we could use it to check the performance of camera lenses.