Slim-CNN: A Light-Weight CNN for Face Attribute Prediction

TL;DR

This paper presents Slim-CNN, a lightweight and efficient neural network architecture for face attribute prediction that maintains high accuracy with significantly fewer parameters, suitable for mobile devices.

Contribution

Introduction of Slim Modules combining depthwise and pointwise convolutions to create a compact, high-accuracy CNN for face attribute prediction.

Findings

Achieves 91.24% accuracy on CelebA dataset.

Uses at least 25 times fewer parameters than comparable methods.

Reduces memory storage by at least 87%.

Abstract

We introduce a computationally-efficient CNN micro-architecture Slim Module to design a lightweight deep neural network Slim-Net for face attribute prediction. Slim Modules are constructed by assembling depthwise separable convolutions with pointwise convolution to produce a computationally efficient module. The problem of facial attribute prediction is challenging because of the large variations in pose, background, illumination, and dataset imbalance. We stack these Slim Modules to devise a compact CNN which still maintains very high accuracy. Additionally, the neural network has a very low memory footprint which makes it suitable for mobile and embedded applications. Experiments on the CelebA dataset show that Slim-Net achieves an accuracy of 91.24% with at least 25 times fewer parameters than comparably performing methods, which reduces the memory storage requirement of Slim-net by at least 87%.

Tables

Table 1. Table 1: Configuration Details of the proposed Slim-Net

Layer Name	Filter Size/ Stride	Squeeze	Expand (1x1)	Expand (3x3)	DW Separable (3x3)
Input Layer of size 178 x 218 x 3
Conv2D (96 filters)	7x7/ 2x2
MaxPool	3x3/2x2
Slim Module		16	64	64	48
MaxPool	3x3/2x2
Slim Module		32	128	128	96
MaxPool	3x3/2x2
Slim Module		48	192	192	144
MaxPool	3x3/2x2
Slim Module		64	256	256	192
MaxPool	3x3/2x2
Global Average Pooling
Fully-Connected	40

Table 2. Table 2: Accuracy, memory and parameter count comparison with state-of-the-art methods. Memory usage reported here is on-disk space used by the model.

Attributes	Liu [Liu et al.(2015)Liu, Luo, Wang, and Tang]	MOON [Rudd et al.(2016)Rudd, Günther, and Boult]	Hand [Hand and Chellappa(2017)]	DMTL [Han et al.(2018)Han, Jain, Wang, Shan, and Chen]	AFFACT [Günther et al.(2016)Günther, Rozsa, and Boult]	PSE [Zhang et al.(2018b)Zhang, Shen, Sun, and Li]	Slim (this work)
5 O’Clock Shadow	91	94.03	94.51	95	94.87	94.85	94.75
Arched Eyebrows	79	82.26	83.42	86	84.21	83.75	83.2
Attractive	81	81.67	83.06	85	82.86	82.95	82.85
Bags Under Eyes	79	84.92	84.92	99	85.31	85.35	85.33
Bald	98	98.77	98.9	99	99.09	98.99	98.84
Bangs	95	95.8	96.05	96	96.1	95.54	96.23
Big Lips	68	71.48	71.47	88	72.7	71.9	71.44
Big Nose	78	84	84.53	92	84.38	84.09	84.28
Black Hair	88	89.4	89.78	85	90.36	89.76	90.13
Blond Hair	95	95.86	96.01	91	96.2	95.5	95.55
Blurry	84	95.67	96.17	96	96.37	96.29	95.42
Brown Hair	80	89.38	89.15	96	88.4	88.66	88.5
Bushy Eyebrows	90	92.62	92.84	85	92.46	91.77	92.79
Chubby	91	95.44	95.67	97	95.74	95.94	95.57
Double Chin	92	96.32	96.32	99	96.52	96.45	96.3
Eyeglasses	99	99.47	99.63	99	99.61	99.65	99.64
Goatee	95	97.04	97.24	98	97.5	97.15	97.55
Gray Hair	97	98.1	98.2	96	98.31	98.23	97.76
Heavy Makeup	90	90.99	91.55	92	92.1	90.39	91.71
High Cheekbones	88	87.01	87.58	88	87.77	87.12	87.8
Male	98	98.1	98.17	98	98.5	98.21	98.12
Mouth Slightly Open	92	93.54	93.74	94	94.06	93.85	93.79
Mustache	95	96.82	96.88	97	97.11	96.85	97
Narrow Eyes	81	86.52	87.23	90	87.69	87.35	87.7
No Beard	95	95.58	96.05	97	96.45	96.22	96.26
Oval Face	66	75.73	75.84	78	77.41	73.98	75.43
Pale Skin	91	97	97.05	97	97.05	97.04	97.18
Pointy Nose	72	76.46	77.47	78	76.9	77.3	77.13
Receding Hairline	89	93.56	93.81	94	93.67	93.55	93.1
Rosy Cheeks	90	94.82	95.16	96	95.16	95.08	95.1
Sideburns	96	97.59	97.85	98	97.84	97.87	97.73
Smiling	92	92.6	92.73	94	92.96	91	92.93
Straight Hair	73	82.26	83.58	85	85.26	84.93	83.14
Wavy Hair	80	82.47	83.91	87	86.25	85.35	83.77
Wearing Earrings	82	89.6	90.43	90	90.98	90.26	90.38
Wearing Hat	99	98.95	99.05	99	99.1	99	99.07
Wearing Lipstick	93	93.93	94.11	93	93.96	93.41	94.18
Wearing Necklace	71	87.04	86.63	89	89.27	88.01	87.12
Wearing Necktie	93	96.63	96.51	97	97.29	97.03	96.7
Young	87	88.08	88.48	90	88.98	88.67	88.14
Average	87	90.94	91.27	92.1	91.67	91.23	91.24
Memory	-	457MB	63.8MB	260MB	98.2MB	-	7.9MB
#Parameters	>100M	136M	15M	65M	26M	62M	0.6M

Table 3. Table 3: Comparing Performance of Slim-Net with other micro-architectures based DNN

Models	Parameters	Memory	Accuracy
Fire Module[Iandola et al.(2016)Iandola, Moskewicz, Ashraf, Han, Dally, and Keutzer]	764,968	9.8MB	80.10%
Inception	1,224,288	15.5MB	90.62%
Residual Block	1,905,096	23.5MB	90.67%
Slim Module	570,600	7.5MB	91.24%

Table 4. Table 4: Effect of varying the kernel size of the layer after SSE Block.

Kernel Size	3x3	5x5	7x7	3x3+3x3
Accuracy	91.24%	91.18%	91.16%	91.20%
Parameters	570,600	591,080	615,848	680,320

Equations

$${\mathcal{L}}_{attr}=-(y)\log (p)+(1-y)\log (1-p)$$

Full text

1 Introduction

The introduction of Convolutional Neural Networks (CNN) has led to the advancement of state-of-the-art performance on several fundamental computer vision tasks such as image classification [Szegedy et al.(2015)Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, Erhan, Vanhoucke, and Rabinovich], semantic segmentation [Chen et al.(2016)Chen, Papandreou, Kokkinos, Murphy, and Yuille], and object detection [Liu et al.(2016)Liu, Anguelov, Erhan, Szegedy, Reed, Fu, and Berg]. The success of popular CNN architectures such as VGG-Net [Simonyan and Zisserman(2014)] and Res-Net [He et al.(2016)He, Zhang, Ren, and Sun] has steered the research into building deeper and more parameter intensive networks. While these networks yield higher accuracy than their predecessors, they are not very economical. This is particularly important when considering deployment on mobile-powered platforms, where the storage burden and computational cost of parameter intensive networks are most noticeable. In recent years, a few efficient architectures, like Mobile-Net [Howard et al.(2017)Howard, Zhu, Chen, Kalenichenko, Wang, Weyand, Andreetto, and Adam] and Shuffle-Net [Zhang et al.(2018a)Zhang, Zhou, Lin, and Sun], have been proposed to address these concerns. Their key idea is to design a deep network using computationally economical types of convolution such as depthwise separable convolution and group convolutions.

Analysis of facial attributes provides a very important visual cue for a range of applications like surveillance, biometric recognition systems, retrieval, fraud detection, etc. Over the last few years, there has been significant interest in the multi-label face attributes problem as shown in Fig.1. Most of the previous work on facial analysis have used Alex-Net [Krizhevsky et al.(2014)Krizhevsky, Sutskever, and Hinton] inspired architectures that made the final network architecture more computationally expensive and parameter exhaustive. This makes it unsuitable for mobile or embedded applications where face attribute prediction is used as one of the main building blocks.

The aim of our work is to design a computationally efficient, high performance CNN architecture that is lighter parameter wise than existing architectures. To achieve this goal we focus on the use of CNN micro-architectures, that is the arrangement of convolutional, pooling, fully-connected, and dropout layers into individual modules. Micro-architectures, like Inception module[Szegedy et al.(2015)Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, Erhan, Vanhoucke, and Rabinovich] and Residual blocks[He et al.(2016)He, Zhang, Ren, and Sun], have been shown to give good discriminative feature representations. In this paper, we propose to use our novel micro-architecture Slim Module to build an efficient deep neural network. The Slim Module is composed of convolutional layers of two different types (depth separable convolution for 3x3 kernel and pointwise convolution for 1x1 kernel) along with skip-connections.

To summarize, our main contributions are as follows:

• •

We introduce a novel CNN micro-architecture Slim Module that is compact, computationally efficient, has a smaller memory footprint, and still yields very good discriminative features.

• •

We thoroughly investigate our proposed micro-architecture for different values of kernel sizes and compare it with other micro-architectures in the literature.

• •

We compare the performance of a stacked Slim Module deep neural network Slim-CNN with state-of-the-art methods on face attribute prediction in terms of accuracy, parameter size, and on-disk memory space.

2 Related Work

We will review two lines of related work: face attribute prediction and micro-architectures in CNNs.

2.1 Face Attribute Prediction

Kumar, et al. [Kumar et al.(2009)Kumar, Berg, Belhumeur, and Nayar] were the first to introduce the classification of facial attributes for the task of face verification. They used binary classifiers to detect the presence or absence of attributes which they called attribute classifiers. Liu, et al. [Liu et al.(2015)Liu, Luo, Wang, and Tang] introduced the large scale CelebFaces Attribute (CelebA) Dataset for prediction of facial attributes in the wild. They proposed a deep learning framework of cascading DNNs consisting of two localization networks (LNet) and an attribute recognition network (ANet) for face localization and face attribute classification. Zhong, et al. [Zhong et al.(2016)Zhong, Sullivan, and Li] used mid-level CNN features to construct deep hierarchical feature representations for facial attribute prediction. The authors argue that some attributes are more locally-defined and earlier layers of CNNs are more suitable for the identification of these attributes. In [Rudd et al.(2016)Rudd, Günther, and Boult], Rudd uses a multi-task approach to solve the problem of attribute classification. In order to deal with the multi-label imbalance problem in the dataset, they introduce mixed objective optimization network(MOON). Since the learning of each facial attribute is treated as a different task, the cost function is set-up for joint optimization.

Hand [Hand and Chellappa(2017)] leveraged the relationships between attributes to build a multi-task deep convolutional network (MCNN) wherein they manually divided 40 attributes into 9 groups according to each attribute’s location. The lower layers of MCNN are shared among all the groups while the higher layers are shared only for related attributes. They attached a fully-connected layer to the output of the MCNN (MCNN-AUX) to identify relationships between different attribute scores and to improve the final accuracy. Han et al. [Han et al.(2018)Han, Jain, Wang, Shan, and Chen] used a multi-task approach to estimate facial attributes by modeling both attribute correlation and attribute heterogeneity in a single network. The authors initialized the network by training it for face verification on the CASIA-Webface dataset [Yi et al.(2014)Yi, Lei, Liao, and Li]. Gunther [Günther et al.(2016)Günther, Rozsa, and Boult] suggested using data augmentation to learn facial attributes on a ResNet architecture without the need of any sort of facial alignment. Note that all of the mentioned works on facial attribute prediction have focused primarily on performance without addressing the need for compact and efficient models.

2.2 Micro-Architecture in CNN

Micro-architecture refers to the organization of individual layers and modules. Over the last few years, micro-architectures such as Inception Modules and Residual Modules have become very popular as building blocks for constructing deep neural networks. Lin [Lin et al.(2013)Lin, Chen, and Yan] proposed a DNN "Network In Network" which consists of stacking multiple micro neural network structures "mlpconv layer" to enhance the abstraction ability of the model. In [Szegedy et al.(2015)Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, Erhan, Vanhoucke, and Rabinovich], Szegedy stacked these Inception Modules to get superior performance while still keeping the computational complexity in check. They implemented dimension reduction by using 1x1 convolution kernels in order to lower the number of parameters. Landola [Iandola et al.(2016)Iandola, Moskewicz, Ashraf, Han, Dally, and Keutzer] presented a compact Fire Module to design a small deep network to get same performance as Alex-Net, but with 50 times fewer parameters. In [He et al.(2016)He, Zhang, Ren, and Sun], He introduced a residual learning framework for training DNNs with very large depths without sustaining a loss of performance. The idea behind He’s work is to fit a residual mapping with a reference to the input rather than to the desired underlying mapping. Several works [He et al.(2017)He, Gkioxari, Dollár, and Girshick, Xie et al.(2017)Xie, Girshick, Dollár, Tu, and He] have since used this principle to build smaller residual modules for training different tasks.

3 Method

Our goal is to design a new CNN micro-architecture to be used as the building blocks of our deep neural network. There has been extensive research into neural network architectures, which has led to many structural innovations. These innovations have contributed significantly toward improving the performance of CNNs across many different computer vision tasks. As such, they were considered when designing our micro-architecture. The following structural innovations have been incorporated into our Slim Module:

1. Multiple Branches: From the success of Inception and ResNets, it can be inferred that having multiple branches in the network provides better feature representations. Multiple branches in a network means multiple paths through the network which would lead to good discriminative features.

2. Small Kernels: VGG-Net [Simonyan and Zisserman(2014)] and GoogleNet [Szegedy et al.(2015)Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, Erhan, Vanhoucke, and Rabinovich] showed that smaller kernels such as [1x1, 3x3, 5x5] can improve the performance of a deep neural network while also reducing the number of parameters of the network. Smaller kernels have the added advantage of being more computationally efficient. Also, stacking two 3x3 convolutional layers have effectively the same receptive field as that of 5x5 convolutional layer. This means that combination of smaller kernels can simulate a larger kernel while keeping the number of parameters significantly lower.

3. Skip-Connections: Skip-Connections [Srivastava et al.(2015)Srivastava, Greff, and Schmidhuber, He et al.(2016)He, Zhang, Ren, and Sun], also called by-pass connections, are defined as the connections between nodes in different layers of a neural network that skip one or more layers of processing. Skip-connections allow neural networks to regulate information by providing paths along which information can go through several layers without any attenuation. This has been highlighted by the success of architectures that utilize skip-connections like in ResNets and HighWay [Srivastava et al.(2015)Srivastava, Greff, and Schmidhuber] Networks.

3.1 Proposed Micro-Architecture:

CNN micro-architecture, as higher level building blocks, present a modular approach to designing deep neural networks by reducing the need to manually choose the type and filter dimensions of each convolutional layer. The proposed Slim module, shown in Figure 2, is made up of Separable Squeeze-Expand (SSE) blocks and depthwise separable convolutional layers. Depthwise separable convolutions consists of depthwise convolution, a spatial convolution performed independently over each channel yielding a new channel space, followed by pointwise convolution, a 1x1 convolution over the channel-space output of the depthwise convolution. These convolutions have fewer parameters than the standard convolutions which make them less prone to overfitting. Additionally, these convolutions are computationally cheaper and faster than the standard convolutions which make them a great fit for mobile and embedded applications.

The SSE block is a multi-layered arrangement consisting of two 1x1 pointwise convolutional layers and a single 3x3 depthwise separable convolutional layer. The first layer, called the "Squeeze" layer, is a 1x1 convolutional layer with fewer filter dimensions than the previous layer which compresses the feature representation. The next layer is a concatenation of a 1x1 pointwise convolutional filter and a 3x3 depthwise separable convolution filter. This is the "Expand" layer as it significantly widens the number of output channels, increasing the filter dimensions to four times that of the squeeze layer. The SSE block is similar to the fire module described in the paper[Iandola et al.(2016)Iandola, Moskewicz, Ashraf, Han, Dally, and Keutzer], but we improve on it by using depthwise separable convolution.

The Slim Module stacks two SSE blocks together followed by a 3x3 depthwise separable convolution layer. Additionally, there is a skip-connection over the first SSE block, resulting in the sum of the input and output of the first block being used as the input to the second SSE block. In our micro-architecture design, the convolutional layers are followed by a batch normalization layer and a ReLu activation layer.

In addition to making the Slim Module cost-effective, we sought to make it easy to configure. To that end, we define a hyperparameter for the Slim Module called FilterCount which is the number of filters in the squeeze layer of the SSE Blocks for a given Slim Module. The number of filters for the other convolutional kernels in that module are just multiples of the FilterCount. To get the number of convolutional filters in the Expand layer and the 3x3 depthwise separable convolutional layer after the second SSE Block, the FilterCount was multiplied by 4 and 3, respectively.

3.2 Network

We use Slim Modules as the building blocks of our deep neural network Slim-CNN. The proposed micro-architecture yields better discriminative features, while reducing the number of required parameters due to the presence of the SSE blocks within the micro-architecture. Four Slim Modules are stacked together to construct the feature extractor portion of the network. Each Slim Module is followed by max-pooling layer. The feature extractor is followed by global average pooling layer whose output is shared by a fully-connected classification layer of size 40 (.i.e. the total number of attributes). A Global Average Pooling layer was compared to fully-connected layers of size 512 and 1024. The Global Average Layer is preferred as it both reduces the number of parameters and improves the performance of the network. In this work, the FilterCount values are set, in order, to [16, 32, 48, 64] for the four Slim Module comprising the Slim-CNN as shown in Figure 3. In this manner, we can design the complete network without exhaustively trying different configurations for each convolutional layer. The design details of Slim-Net is shown in Table 1.

Since the task at hand is a binary, multi-label classification problem, we use a sigmoid cross-entropy loss function for each attribute during training. The loss function for each attribute is defined as:

[TABLE]

where y is the attribute label and p is the prediction from corresponding attribute label.

4 Experiments and Results

4.1 Implementation Details

All experiments in this work were run using GeForce GTX 1080 Ti on Intel Core i7 with 31 GiB of memory using the Keras Python Deep Learning Library with a TensorFlow backend. For training the network, we used the ADAM optimizer with a learning rate of 0.0001. The network was trained with glorot initialization [Glorot and Bengio(2010)]. The batch-size is set to 64 and the depthwise separable convolution layers in the Slim modules apply L2 regularization(0.0001).

4.2 Dataset

In this work, we use the CelebA dataset, which consists of over 200,000 celebrity images where each image has been annotated with 40 different attributes. The dataset covers large variations in pose, background, and illumination which makes predicting attributes extremely challenging. The images in the dataset are split in the following fashion: 160,000 images for training, 20,000 images for validation, and 20,000 images for final testing. Some of the images from the CelebA dataset are shown in Fig 1. The CelebA database is an imbalanced set which further illustrates the difficulty of the task.

4.3 Attribute Prediction Results

In this section, we describe the results of many experiments comparing our Slim-CNN against other methods and micro-architectures.

4.3.1 Comparison with other methods

Table 2 shows how our Slim-Module inspired deep neural network performs as compared to other methods in terms of both accuracy for each of the 40 attributes and overall model size. As can be seen, our Slim-CNN is very competitive with state-of-the-art methods, while only using a fraction of the parameters of its competitors. Some of these methods use computationally expensive network architecture as the backbone of their proposed solutions. For example, PSE [Zhang et al.(2018b)Zhang, Shen, Sun, and Li], DMTL[Han et al.(2018)Han, Jain, Wang, Shan, and Chen], and Liu [Liu et al.(2015)Liu, Luo, Wang, and Tang] are inspired by the parameter intensive Alex-Net (60M parameters) while AFFACT [Günther et al.(2016)Günther, Rozsa, and Boult] is built on top of a ResNet-50 architecture with more than 25M parameters. The reduction in number of parameters of our proposed network is staggering, making it ideal for mobile and embedded applications. We report the on-disk memory needed by the networks for several methods. The proposed Slim-Net is extremely light and require significantly less memory space as compared to other methods, while still achieving high accuracy.

4.3.2 Comparison of Slim-CNN with other Micro-Architecture-Based DNNs

Here, we compare our Slim-CNN with other micro-architecture-based DNNs such as Squeeze-Net. To construct the Squeeze-Net DNN used in our experiments, we removed the 3x3 depthwise separable convolution layer after the second SSE Block and the skip connection in each Slim Module. The filter sizes for all the convolutional layers in the resulting Fire modules remain the same as those of the Slim Module. We also construct two different CNNs for comparing the Slim Module with inception modules and residual blocks. These CNNs were constructed by replacing Slim Modules with their corresponding micro-architectures i.e. inception modules and residual blocks. For a fair comparison, we try to keep the configurations of these CNNs consistent with Slim-CNN. For the inception inspired CNN, the number of filters in the 3x3 kernels of the inception module is the same as the number of filters in the 3x3 kernel in the expand layer of SSE blocks of the Slim Module. Similarly for the residual block CNN, the number of filters for the 1x1 and 3x3 kernels in the residual block is twice that of the 1x1 kernel in the squeeze layer and the 3x3 kernel in the expand layer of SSE Block. We tested different configurations for these two CNNs and reported the best results here. We provide more details on the configurations for residual block CNN and inception module CNN in the supplementary material. Table 3 shows that Slim-CNN performs better than inception CNN and residual CNN despite having fewer parameters than both. This shows that the Slim Module can efficiently generate good feature representation for faces. Slim-Net improves on the Fire module[Iandola et al.(2016)Iandola, Moskewicz, Ashraf, Han, Dally, and Keutzer] based DNN results by over 11% which shows that the addition of a 3x3 kernel after the SSE Blocks and the skip-connection to micro-architecture yields better feature representations.

4.3.3 Varying the kernel size of the layer after SSE Block

The Slim Module uses 3x3 depthwise separable convolution layer after the two SSE Blocks. We opted to use a kernel of this size for the final layer of the Slim Module after running several experiments on different kernel sizes. The final layer configurations that were tested are: (i) a single 3x3 depthwise separable convolution layer (ii) a single 5x5 depthwise separable convolution layer (iii) a single 7x7 depthwise separable convolution layer and (iv) two 3x3 depthwise separable convolution layers in series. Table: 4 shows nearly identical performance from each configuration, suggesting that increasing the size of kernels does not necessarily increase performance. As such, the single 3x3 kernel, having the fewest parameters, was chosen as the configuration for the final layer of the module. This choice was made to be consistent with our design goal of minimizing the parameter space.

5 Conclusion

We proposed an efficient micro-architecture Slim Module for designing deep neural networks to be used in mobile and embedded applications. Our Slim-CNN has shown comparable performance to the state-of-the-art methods while significantly reducing the number of parameters. Specifically, Slim-CNN achieves an average accuracy of 91.24% with only 600K parameters. The next smallest state-of-the-art architecture [Hand and Chellappa(2017)] has more than 25 times the number of parameters at 16M. We verify the effectiveness of the Slim Module for the face attribute task by comparing it with other well-known micro-architectures in literature. The Slim Module micro-architecture outperforms the Inception module, Residual block, and Fire-Module based SqueezeNet [Iandola et al.(2016)Iandola, Moskewicz, Ashraf, Han, Dally, and Keutzer]. Moreover, the incredibly low parameter footprint of Slim-CNN makes it very well suited to resource limited environments.