Citation Context

...er The outputs of Gabor filter are regarded as sparse coding, because Gabor-like receptive fields are obtained by using the constraint which maximizes the sparseness of the response to natural images =-=[9]-=-. It is also reported that Gabor-like filters are obtained by using the info-max network which is able to perform independent components analysis [10]. This means that the outputs of Gabor filters are...

Citation Context

...is necessary that one kernel value is computed from local kernels. The product and summation of local kernels are considered as the integration methods of local kernels which satisfy Mercer’s theorem =-=[6, 7]-=-. As described above, how to integrate the local similarities (kernels) is important. It is considered that the local summation kernel is better than local product kernel. The product value of local k...

Citation Context

...base is described in section 3.1. Next, the comparison results and face detection results are shown in section 3.2. 3.1. Image Database In this paper, HOIP face database 1 and CMU test face database =-=[11]-=- are used as face images. The face regions of these images are cropped by using the positions of eyes, nose, and mouth. The number of cropped face images is 933 (HOIP:300 and CMU:633). Since the numbe...

Citation Context

...NTRODUCTION Face detection is the first essential step for automatic face recognition. Since automatic face recognition has many potential applications, face detection becomes an active research area =-=[1, 2]-=-. Some frontal face detection methods give high detection rate under the restricted environment [1, 2]. However, in practical environment and applications, there is some obstacles such as occlusion, i...

Citation Context

...ed from AR face database [15]. The shadow on a face is considered as one of the occlusion. Therefore, the face with shadow is also detected correctly. This image is obtained from Yale face database B =-=[16]-=-. In addition, some faces with occlusion are detected correctly. 4. CONCLUSIONS We present a robust face detector under partial occlusion. The conventional face detection methods based on SVM use glob...

Citation Context

...e shadows on faces induced by illumination direction changes are considered as one of the occlusion. In recent years, the effectiveness of Support Vector Machine (SVM) to object detection is reported =-=[3, 4]-=-. However, in the conventional approaches, one kernel is applied to global features. Therefore, those methods are not robust to occlusion, because global features are influenced easily by noise or occ...

Citation Context

...is necessary that one kernel value is computed from local kernels. The product and summation of local kernels are considered as the integration methods of local kernels which satisfy Mercer’s theorem =-=[6, 7]-=-. As described above, how to integrate the local similarities (kernels) is important. It is considered that the local summation kernel is better than local product kernel. The product value of local k...

Citation Context

...s are integrated well, it is expected that the robustness to occlusion is realized. For example, Martinez realized the robust recognition under partial occlusion by integrating the local similarities =-=[5]-=-. From these considerations, in order to give SVM the robustness under partial occlusion, it is necessary to treat local features in SVM. In that method, the robustness to occlusion is realized by int...

Citation Context

...e face detection results of the proposed method are shown in Figure 4. We understand that the faces wearing sunglasses or scarf are detected correctly. These images are obtained from AR face database =-=[15]-=-. The shadow on a face is considered as one of the occlusion. Therefore, the face with shadow is also detected correctly. This image is obtained from Yale face database B [16]. In addition, some faces...

Citation Context

...n kernel. In this paper, the robust face detector under partial occlusion is realized by using SVM with local summation kernel. In recent years, a face detection method based on local SVM is proposed =-=[8]-=-. Although that method uses local features, SVM is applied to all features extracted from local region. Namely, the global kernel is applied to local region. Therefore, that method is different from t...

Citation Context

...e shadows on faces induced by illumination direction changes are considered as one of the occlusion. In recent years, the effectiveness of Support Vector Machine (SVM) to object detection is reported =-=[3, 4]-=-. However, in the conventional approaches, one kernel is applied to global features. Therefore, those methods are not robust to occlusion, because global features are influenced easily by noise or occ...

Citation Context

...NTRODUCTION Face detection is the first essential step for automatic face recognition. Since automatic face recognition has many potential applications, face detection becomes an active research area =-=[1, 2]-=-. Some frontal face detection methods give high detection rate under the restricted environment [1, 2]. However, in practical environment and applications, there is some obstacles such as occlusion, i...

Citation Context

...s the sparseness of the response to natural images [9]. It is also reported that Gabor-like filters are obtained by using the info-max network which is able to perform independent components analysis =-=[10]-=-. This means that the outputs of Gabor filters are regarded as independent of each other. Gabor filters are defined by ψk (x) = k2 � 2 2 −k x exp σ2 2σ2 � � � �� 2 exp (ikx) − exp −σ /2 , (1) where x ...

Citation Context

...d 5 times by this processing. The shifting of face images is performed only to the first and second face sets. On the other hand, the non-face images are obtained by PICS database [12], pbic database =-=[13]-=-, and WWW. The 17, 750 images with 38 × 38 pixels are cropped randomly from PICS and WWW images. Examples of non-face images are shown in Figure 1 (b). These images are divided into 2 sets. The first ...