Image Spam Detection with VGG19

This is a model for detecting spam in images using transfer learning, with a VGG19 base model fine-tuned on the Image Spam Hunter Dataset.

Model Description

This model is built for binary classification, detecting whether an image is spam or not. The VGG19 architecture was used as the base model for transfer learning, leveraging pre-trained weights from ImageNet. The final layers were customised to include a GlobalAveragePooling layer, followed by fully connected layers, and a softmax activation for classification.

Base Model: VGG19
Task: Binary classification (Spam / Not Spam)
Framework: Keras (TensorFlow backend)

Training and Evaluation Data

Dataset

The model was trained and evaluated on the Image Spam Hunter Dataset, a publicly available dataset that contains spam and non-spam images:

Training set: 1731 images
Validation set: 1385 images

You can find more about the dataset here.

Training Procedure

Preprocessing

Images were resized to 256x256.
Normalisation was applied to scale pixel values between 0 and 1.

Optimizer

Optimizer: Adam
Learning Rate: 0.001

Training Parameters

Batch Size: 128
Epochs: 26 (early stopping)

Loss Function

Binary Cross-Entropy Loss

Augmentations

Random augmentation using Keras CV RandAugment layer.

Training Metrics

Below is a summary of the model's performance over the training process:

	FN	FP	TN	TP	accuracy	loss	val_FN	val_FP	val_TN	val_TP	val_accuracy	val_loss
0	187	140	499	559	0.763899	1.93435	19	2	169	156	0.939306	0.378412
1	16	35	604	730	0.963177	0.242954	6	3	168	169	0.973988	0.186523
2	29	19	620	717	0.965343	0.174301	3	12	159	172	0.956647	0.149353
3	17	26	613	729	0.968953	0.116907	3	7	164	172	0.971098	0.0973678
4	14	17	622	732	0.977617	0.0807728	2	12	159	173	0.959538	0.0991694
5	13	18	621	733	0.977617	0.0713564	3	6	165	172	0.973988	0.0604228
6	10	17	622	736	0.980505	0.0604733	3	5	166	172	0.976879	0.0584633
7	9	13	626	737	0.984116	0.0478682	4	4	167	171	0.976879	0.0745383
8	5	13	626	741	0.987004	0.0376507	3	7	164	172	0.971098	0.0773135
9	7	6	633	739	0.990614	0.0248621	3	2	169	172	0.985549	0.0628484
10	7	10	629	739	0.987726	0.0298408	2	2	169	173	0.988439	0.0531744
11	4	10	629	742	0.989892	0.0242986	3	2	169	172	0.985549	0.0675473
12	6	7	632	740	0.990614	0.0326418	2	3	168	173	0.985549	0.0671533
13	6	9	630	740	0.98917	0.0279278	2	5	166	173	0.979769	0.06897
14	7	7	632	739	0.989892	0.0244297	2	3	168	173	0.985549	0.0550619
15	5	4	635	741	0.993502	0.0228877	3	2	169	172	0.985549	0.0558347
16	4	10	629	742	0.989892	0.0277857	4	0	171	171	0.988439	0.0401906
17	2	7	632	744	0.993502	0.0193213	9	1	170	166	0.971098	0.0959367
18	7	3	636	739	0.99278	0.0151094	4	6	165	171	0.971098	0.0873069
19	7	3	636	739	0.99278	0.017344	1	6	165	174	0.979769	0.0964724
20	5	4	635	741	0.993502	0.024251	1	4	167	174	0.985549	0.0580115
21	3	9	630	743	0.991336	0.0207464	1	3	168	174	0.988439	0.0510422
22	2	6	633	744	0.994224	0.0122992	1	1	170	174	0.99422	0.0439666
23	3	2	637	743	0.99639	0.0105987	4	2	169	171	0.982659	0.0502425
24	3	6	633	743	0.993502	0.0164827	6	1	170	169	0.979769	0.0865423
25	5	7	632	741	0.991336	0.0245302	6	3	168	169	0.973988	0.0918282
26	9	8	631	737	0.987726	0.0296158	5	1	170	170	0.982659	0.0991676

Model Plot

The model architecture can be visualised in the image below:

Validation Accuracy

The final validation accuracy achieved by the model was 98%.

How to Use the Model

Loading the Model

To load this model in your application, use the following:

from tensorflow.keras.models import load_model
model = load_model("path/to/saved_model")

fbadine
/

image-spam-detection