Page 84 - 2024-Vol20-Issue2

80 | AlKarem, Khalid, & Ali

conv2d_3_input input: [(None,224,224,1)] TABLE I.
Input Layer output: [(None,224,224,1)] THE PROPOSED MODEL ARCHITECTURE

conv2d_3 input: (None,224,224,1) Layer Input Setting
Conv2D output: (None,222,222,32) Input layer 64×64×1
Convolution 32×3× 3
max_pooling2d-3 input: (None,222,222,32) Max pooling
Maxpooling2D output: (None,111,111,32) Convolution 2×2
Max pooling 64 ×3 × 3
conv2d_4 input: (None,111,111,32) Convolution
Conv2D output: (None,109,109,64) Max pooling 2×2
Flatten layer 128 ×3 × 3
max_pooling2d-4 input: (None,109,109,64)
Max_pooling2D output: (None,54,54,64) Fully connected layer 2×2

Conv2d_5 input: (None,54,54,64) Dropout 128 neurons, activation
Conv2D output: (None,52,52,128) function : Relu
Fully connected layer
0.2
2 classes , activation function
: Softmax classifier

max_pooling2d-5 input: (None,52,52,128) ”sparse categorical crossentropy” as loos function to compute
Max_pooling2D output: (None,26,26,128) the error between given results and predicated results during
training. Model fitting is performed based on the specified
flatten_1 input: (None,26,26,128) data of training and validation utilizing batch size of 32, and
Flatten output: (None,86528) the no of epoch is 15.

dense_2 input: (None,86528) III. RESULTS AND DISCUSSION
Dense output: (None,128)
This paper is implemented a deep learning method for of-
dropout_1 input: (None,128) fline handwritten signature recognition via introducing a deep
Dropout output: (None,128) architecture model based on CNN network. The method im-
plemented using programming language instructions of Keras
dense_3 input: (None,128) Tensorflow library of Python. GPU of Google Colab platform
Dense output: (None,2) is utilized to run the code on HP laptop with Intel core i7,
and Nvidia Getforce GPU. This study executed a method to
Fig. 2. Plot of the proposed model architecture. recognize genuine signature from forgery signature utilizing
collected dataset of scanned images of handwritten signature
Fig. 2. Plot of the proposed model architecture. taken from many individuals. The collected dataset consists
of 720 images, 360 images for genuine signature, and 360
3×3; while the number of filters for three layers is 32, 64, and images for forgery signatures, to be used for training and vali-
128 respectively. Max pooling layer is proposed after each dating the proposed model. The model is implemented after
convolution layer to keep good features as depicted in model randomly splitting the data into train, validate, and test images.
architecture Table I. This work implemented a method based on CNN network for
features extraction procedure while dense network or deep
E. Model Training and Fitting neural network is proposed for features classification process
For model training, the features are first flattened into features as illustrated in section II. This study conducted a quantita-
vector, and then fully connected layer (Dense network) is pro- tive measurement based on accuracy metric to evaluate the
posed for features processing and recognition, which assigned performance of this method. The proposed model based on
128 neurons, by using activation function Relu. Dropout CNN network utilized training accuracy, training loss error,
layer is suggested with rate 0.2 to avoid overfitting, due to validation accuracy, and validation loos error to evaluate the
dropout can skip some neurons every time to avoid too much functionality of the model during fitting process. The batch
training and by that overfitting problem could be overcome. size is 32 and epoch no is 15. The acquired accuracy for
Then dense layer with Softmax activation function is pro- validation dataset is 0.9536 %. The outcomes of model train-
posed to classify the features into two classes. The opti-
mization function that has been proposed for model training
is “Adam” with learning rate 0.001. This work employed