Search

The identification of textile fiber materials is a tedious task. Traditional physical or chemical methods rely on specialist knowledge and expensive instruments. Currently, the deep learning-based methods are limited by the lack of sufficient image samples. In this paper, we propose a method for the classification of textile materials based on advanced densely connected convolutional networks (Densenet) using fabric surface images. Firstly, we selected a small sample dataset containing five types of materials and performed some pre-processing, including colour weakening and data enhancement. The lightweight network Densenet was then used as the main body to ensure the low parameters of the network, making it more suitable for the recognition of small sample image datasets. The network is then fused with the attention mechanism, Orthogonal Softmax Layer (OSL) and the depth-separable convolution mechanism, and we call the final network AONet. It has only 0.64 M (million) parameters. In this paper, we tested four metrics of AONet for five types of fabric images, and the results showed that the recognition accuracy was 95.61%. The precision, recall and F1 scores were 94.87%, 94.31% and 94.40%, respectively. [ABSTRACT FROM AUTHOR]

Current researches have demonstrated that illumination variation on face images degrades the accuracy of facial identity and emotion recognition. To decrease the impact of illumination variation, researchers have proposed many creative methods of illumination compensation. However, these methods are limited in compensating for the shadow around the nose. On the basis of our previous researches, we now propose a novel approach which can effectively decrease the impact of illumination variation, especially the shadow around the nose. Firstly, we preprocessed the face image with uneven brightness using technologies of illuminant direction estimation and improved Retinex. Secondly, we turn the original face image into a binary image with only shadow region or non-shadow region using region growing technology. Thirdly, we calculate the difference between the intensity of the original input face image and the average intensity of the face images under the frontal illumination. Fourthly, for the face image preprocessed in the first step, we keep its non-shadow region. For the intensity difference, we extract its shadow region whose intensity is reduced by an adaptive value. Fifthly, we synthesize the non-shadow region and the shadow region in step four. Finally, we apply maximum filter to smooth the boundary between them. The proposed method is simple in computation and does not need any training steps or any knowledge of 3D models. The experimental results using extended Yale face database B show that our method achieves better illumination compensation comparing with the existing techniques, and provide more satisfactory experimental data for facial identity and emotion recognition. [ABSTRACT FROM AUTHOR]

Illumination normalization of face image for face recognition and facial expression recognition is one of the most frequent and difficult problems in image processing. In order to obtain a face image with normal illumination, our method firstly divides the input face image into sixteen local regions and calculates the edge level percentage in each of them. Secondly, three local regions, which meet the requirements of lower complexity and larger average gray value, are selected to calculate the final illuminant direction according to the error function between the measured intensity and the calculated intensity, and the constraint function for an infinite light source model. After knowing the final illuminant direction of the input face image, the Retinex algorithm is improved from two aspects: (1) we optimize the surround function; (2) we intercept the values in both ends of histogram of face image, determine the range of gray levels, and stretch the range of gray levels into the dynamic range of display device. Finally, we achieve illumination normalization and get the final face image. Unlike previous illumination normalization approaches, the method proposed in this paper does not require any training step or any knowledge of 3D face and reflective surface model. The experimental results using extended Yale face database B and CMU-PIE show that our method achieves better normalization effect comparing with the existing techniques. [ABSTRACT FROM AUTHOR]