Method and system for fashion attribute detection
Traditional systems used for fashion attribute detection struggle to generate accurate predictions due to presence of large intra-class and relatively small inter-class variations in data related to the fashion attributes. The disclosure herein generally relates to image processing, and, more particularly, to a method and system for fashion attribute detection. The method proposes F-AttNet, an attribute extraction network to leverage the performance of fine-grained localized fashion attribute recognition. F-AttNet comprises Attentive Multi-scale Feature Encoder (AMF) blocks that encapsulate multi-scale fine-grained attribute information upon adaptive recalibration of channel weights. F-AttNet is designed by hierarchically stacking the AMF encoders to extract deep fine-grained information across multiple scales. A data model used by F-AttNet is trained using a novel γ-variant focal loss function for addressing the class imbalance problem by penalizing wrongly classified examples and incorporating separate importance to positive and negative instances.
1. A processor implemented method of fashion feature extraction, comprising:
collecting at least one image as input, via one or more hardware processors; and
processing the at least one image using a feature extraction network comprising a plurality of Attentive Multi-scale Feature (AMF) blocks implemented the via one or more hardware processors, using a data model, wherein processing the at least one image by the plurality of AMF blocks comprising:
extracting a plurality of features from the at least one image, by a first subnetwork of the AMF blocks, wherein the first subnetwork enables extraction of coarse features in parallel manner to aggregate different representations from low-level features for fine-grained image analysis;
identifying and extracting features belonging to different scales, from among the plurality of features extracted from the at least one image, by a second subnetwork of the AMF blocks, wherein the second subnetwork applies a convolution operation on the plurality of features, wherein extracting the plurality of features from the at least one image comprises concatenating a plurality of feature representations obtained from the at least one image by applying the convolution operation on the at least one image;
assigning a unique weightage to each of a plurality of channels used for the convolution operation, based on a determined importance of each of the features belonging to the different scales, by a third subnetwork of the AMF blocks for adaptive channel calibration;
determining a rank for each of the extracted features belonging to the different scales, based on the unique weightage of corresponding channel, by the third subnetwork; and
generating one or more recommendations of the extracted features based on the determined rank of each of the extracted features; and
verifying accuracy of the generated one or more recommendations of the extracted features using a γ-variant focal loss function, wherein the γ-variant focal loss function is used to train a data model for attribute extraction for addressing class imbalance by penalizing wrongly classified examples and incorporating importance to positive and negative instances, wherein the γ-variant focal loss function is provided by:
L
(
y
t
,
y
p
)
=
1
C
∑
i
=
1
C
[
-
y
t
(
i
)
[
(
1
-
y
p
(
i
)
γ
1
)
log
(
y
p
(
i
)
)
]
n
-
λ
(
1
-
y
t
(
i
)
)
[
(
y
p
(
i
)
γ
2
)
log
(
1
-
y
p
(
i
)
)
]
n
]
wherein, y t and y p denote ground-truth labels and predicted labels, hyper-parameters γ 1 and γ 2 enable the γ-variant focal loss function to adaptively focus on false positive and false negative hard examples by increasing corresponding cost in the loss function, wherein λ deals with providing different weights to the positive and negative instances, wherein γ 1 and γ 2 are used by the γ-variant focal loss function to separately optimize the attribute extraction network by reducing all the false instances depending on their probability of occurrence for true and false instances.
2. The method of claim 1 , wherein identifying and extracting features belonging to different scales by applying the convolution operation on the plurality of features comprises:
passing the features extracted from the at least one image through a plurality of paths;
encapsulating a plurality of multi scale fine-grained feature maps of the features in each of the plurality of paths, wherein a stride length in each path is adaptively varied, wherein the encapsulation of plurality of multi scale fine-grained feature maps helps to represent the features required for multi-label feature extraction and classification;
obtaining a plurality of fine-tuned features from the multi scale fine-grained feature maps, as an aggregated response to distinguish between visually similar attributes in a multi-label classification; and
obtaining a summation of the aggregated response and the features extracted from the at least one image as the features belonging to the different scales for discrimination of similar-looking attributes.
3. The method of claim 1 , wherein the data model is trained to perform the feature attribute extraction, further comprising:
determining value of a plurality of micro-metrics for a dataset, wherein the plurality of micro-metrics comprises precision, recall, and F1 score;
determining an optimum value of each of a plurality of hyper parameters of a γ-variant focal loss function, based on the determined values of the plurality of micro-metrics; and
training a data model using the determined optimum value of the plurality of hyper parameters of the γ-variant focal loss function.
4. A system for fashion feature extraction, comprising:
one or more hardware processors;
a communication interface; and
a memory storing a plurality of instructions, wherein the plurality of instructions cause the one or more hardware processors to:
collect at least one image as input; and
process the at least one image using a feature extraction network comprising a plurality of Attentive Multi-scale Feature (AMF) blocks implemented the via one or more hardware processors, using a data model, wherein processing the at least one image by the plurality of AMF blocks comprising:
extracting a plurality of features from the at least one image, by a first subnetwork of the AMF blocks, wherein the first subnetwork enables extraction of coarse features in parallel manner to aggregate different representations from low-level features for fine-grained image analysis;
identifying and extracting features belonging to different scales, from among the plurality of features extracted from the at least one image, by a second subnetwork of the AMF blocks, wherein the second subnetwork applies a convolution operation on the plurality of features, wherein extracting the plurality of features from the at least one image comprises concatenating a plurality of feature representations obtained from the at least one image by applying the convolution operation on the at least one image;
assigning a unique weightage to each of a plurality of channels used for the convolution operation, based on a determined importance of each of the features belonging to the different scales, by a third subnetwork of the AMF blocks for adaptive channel calibration;
determining a rank for each of the extracted features belonging to the different scales, based on the unique weightage of corresponding channel, by the third subnetwork; and
generating one or more recommendations of the extracted features based on the determined rank of each of the extracted features, and
verifying accuracy of the generated one or more recommendations of the extracted features using a γ-variant focal loss function, wherein the γ-variant focal loss function is used to train a data model for attribute extraction for addressing class imbalance by penalizing wrongly classified examples and incorporating importance to positive and negative instances, wherein the γ-variant focal loss function is provided by:
L
(
y
t
,
y
p
)
=
1
C
∑
i
=
1
C
[
-
y
t
(
l
)
[
(
1
-
y
p
(
i
)
γ
1
)
log
(
y
p
(
i
)
)
]
n
-
λ
(
1
-
y
t
(
i
)
)
[
(
y
p
(
i
)
γ
2
)
log
(
1
-
y
p
(
i
)
)
]
n
]
wherein, y t and y p denote ground-truth labels and predicted labels, hyper-parameters, γ 1 and γ 2 enable the γ-variant focal loss function to adaptively focus on false positive and false negative hard examples by increasing corresponding cost in the loss function, wherein λ deals with providing different weights to the positive and negative instances, wherein γ 1 and γ 2 are used by the γ-variant focal loss function to separately optimize the attribute extraction network by reducing all the false instances depending on their probability of occurrence for true and false instances.
5. The system of claim 4 , wherein the system identifies and extracts features belonging to different scales by applying the convolution operation on the plurality of features, by:
passing the features extracted from the at least one image through a plurality of paths;
encapsulating a plurality of multi scale fine-grained feature maps of the features in each of the plurality of paths, wherein a stride length in each path is adaptively varied, wherein the encapsulation of plurality of multi scale fine-grained feature maps helps to represent the features required for multi-label feature extraction and classification;
obtaining a plurality of fine-tuned features from the multi scale fine-grained feature maps, as an aggregated response to distinguish between visually similar attributes in a multi-label classification; and
obtaining a summation of the aggregated response and the features extracted from the at least one image as the features belonging to different scales for discrimination of similar-looking attributes.
6. The system of claim 4 , wherein the system trains the data model to perform the feature attribute extraction, by:
determining value of a plurality of metrics for a dataset, wherein the plurality of micro-metrics comprises precision, recall, and F1 score;
determining an optimum value of each of a plurality of hyper parameters of a γ-variant focal loss function, based on the determined values of the plurality of micro-metrics; and
training a data model using the determined optimum value of the plurality of hyper parameters of the γ-variant focal loss function.
7. One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:
collecting at least one image as input, via one or more hardware processors;
processing the at least one image using a feature extraction network comprising a plurality of Attentive Multi-scale Feature (AMF) blocks implemented the using a data model, wherein processing the at least one image by the plurality of AMF blocks comprising:
extracting a plurality of features from the at least one image, by a first subnetwork of the AMF blocks, wherein the first subnetwork enables extraction of coarse features in parallel manner to aggregate different representations from low-level features for fine-grained image analysis;
identifying and extracting features belonging to different scales, from among the plurality of features extracted from the at least one image, by a second subnetwork of the AMF blocks, wherein the second subnetwork applies a convolution operation on the plurality of features, wherein extracting the plurality of features from the at least one image comprises concatenating a plurality of feature representations obtained from the at least one image by applying the convolution operation on the at least one image;
assigning a unique weightage to each of a plurality of channels used for the convolution operation, based on a determined importance of each of the features belonging to the different scales, by a third subnetwork of the AMF blocks for adaptive channel calibration;
determining a rank for each of the extracted features belonging to the different scales, based on the unique weightage of corresponding channel, by the third subnetwork; and
generating one or more recommendations of the extracted features based on the determined rank of each of the extracted features, and
verifying accuracy of the generated one or more recommendations of the extracted features using a γ-variant focal loss function, wherein the γ-variant focal loss function is used to train a data model for attribute extraction for addressing class imbalance by penalizing wrongly classified examples and incorporating importance to positive and negative instances, wherein the γ-variant focal loss function is provided by:
L
(
y
t
,
y
p
)
=
1
C
∑
i
=
1
C
[
-
y
t
(
i
)
[
(
1
-
y
p
(
i
)
γ
1
)
log
(
y
p
(
i
)
]
n
-
λ
(
1
-
y
t
(
i
)
)
[
(
y
p
(
i
)
γ
2
)
log
(
1
-
y
p
(
i
)
)
]
n
]
wherein, y t and y p denote ground-truth labels and predicted labels, hyper-parameters γ 1 and γ 2 enable the γ-variant focal loss function to adaptively focus on false positive and false negative hard examples by increasing corresponding cost in the loss function, wherein λ deals with providing different weights to the positive and negative instances, wherein γ 1 and γ 2 are used by the γ-variant focal loss function to separately optimize the attribute extraction network by reducing all the false instances depending on their probability of occurrence for true and false instances.
8. The one or more non-transitory machine-readable information storage mediums of claim 7 , wherein identifying and extracting features belonging to different scales by applying the convolution operation on the plurality of features comprises:
passing the features extracted from the at least one image through a plurality of paths;
encapsulating a plurality of multi scale fine-grained feature maps of the features in each of the plurality of paths, wherein a stride length in each path is adaptively varied, wherein the encapsulation of plurality of multi scale fine-grained feature maps helps to represent the features required for multi-label feature extraction and classification;
obtaining a plurality of fine-tuned features from the multi scale fine-grained feature maps, as an aggregated response to distinguish between visually similar attributes in a multi-label classification; and
obtaining a summation of the aggregated response and the features extracted from the at least one image as the features belonging to the different scales for discrimination of similar-looking attributes.
9. The one or more non-transitory machine-readable information storage mediums of claim 7 , wherein the data model is trained to perform the feature attribute extraction, comprising:
determining value of a plurality of micro-metrics for a dataset, wherein the plurality of micro-metrics comprises precision, recall, and F1 score;
determining an optimum value of each of a plurality of hyper parameters of a γ-variant focal loss function, based on the determined values of the plurality of micro-metrics; and
training a data model using the determined optimum value of the plurality of hyper parameters of the γ-variant focal loss function.