Pokepedia: Pokemon Image Classification Using Transfer Learning

In recent years, deep learning techniques have gained significant attention in various fields such as Pattern Recognition [1-9], Medical Imaging, Video Analysis, driver drowsiness detection [10, 11], video analysis, Spam detection [12], Healthcare, Clustering [13] and many more. One of the popular techniques is transfer learning, which allows the pre-trained models to be used for a new set of tasks with minimal training data. Transfer learning has shown promising results in image classification tasks, and it can significantly reduce training time and improve model performance.

The classification of Pokémon images can help in building a robust dataset for various applications such as gaming, educational purposes, and research. Due to the complex features and variations in Pokémon images, it is challenging to achieve high accuracy with traditional image classification techniques. Therefore, this research problem can be addressed using transfer learning techniques.

Existing studies on Pokemon image classification using transfer learning have shown promising results in accurately identifying Pokemon species in various scenarios and backgrounds. These studies can be categorized into different areas, such as data augmentation techniques, model architectures, and hyperparameter tuning. Despite the promising results, previous studies have some limitations, such as low accuracy, overfitting, and insufficient dataset size. To overcome these limitations, this report proposes a novel approach for the classification of Pokémon images using transfer learning and data augmentation techniques.

To solve the defects identified in the existing studies, this paper establishes a model based on transfer learning and data augmentation techniques and applies it to classify a dataset of Pokémon images. The aim of this study is twofold: first, it proposes a novel approach that achieves high accuracy in classifying Pokémon images, and second, it demonstrates the effectiveness of transfer learning and data augmentation techniques in improving model performance. The findings shed new light on the effectiveness of transfer learning and data augmentation in improving the performance of deep learning models for image classification tasks, specifically in the context of Pokémon images.

Section 1 presents the research background and motivation followed by a comprehensive review of the existing studies on transfer learning in Section 2. Section 3 describes the methodology utilized in this study, including the datasets, pre-trained models, and evaluation metrics. Section 4 presents the results of the experiments and provides a comparative analysis of the performance of the pre-trained models.  Section 5 conclusion finally, discuss the potential for future work in Section 6.

3.1 Dataset augmentation

In this work, we have created the Pokémon dataset with 600 images collected from internet sources. Pokémon dataset consists of ten types of Pokémon characters that are Pikachu, Raichu, Charmander, Bulbasaur, Squirtle, Eevee, Piplup, Snorlax, Jigglypuff, Psyduck. The sample images in the dataset are shown in Figure 1. Data Augmentation was used to increase the images to 7386 images. In this work, we used the techniques like Scaling, Padding, Cropping, Flipping, Rotation, Translation, and Affine transformation for Data Augmentation.

1.png

Figure 1. Sample images of dataset

3.2 Pokemon characters classification using transfer learning

2.png

Figure 2.  Framework of proposed work

In transfer learning, the knowledge of an already trained model is applied to different but related problems. In this case, the model's output from a layer prior to the output layer is used as input to a new classifier model, which operates the first and middle layers and retrains the following layers. This fine-tuning can have built a solid machine learning model with comparably little data and this method can reduce the training time for a neural network model. Here we used MobileNetV2, EfficientNetB7, EfficientNetV2L, DenseNet201, ResNet101, VGG19, VGG16 pre-trained models to predict and classify the Pokémon images. The framework of the proposed work givenin Figure 2.

3.2.1 VGG 16 and VGG19

Simonyan and Zisserman [20] created VGG16 for the ILSVRC 2014 competition. It includes sixteen convolutional layers with the most effective 3x3 kernels. The layout opted through authors is just like AlexNet i.e., growth in the quantity of the capabilities map or convolution because of the intensity of the community increases. The community incorporates 138 million parameters. In work, we changed the 1000 classes with our 10 classes. Accuracy is obtained using the Adam optimizer. Similarly, by pushing the intensity to 19 layers VGG19 structure is defined.

3.2.2 MobileNetV2

MobileNet-v2 [21] is a family of computer vision models for mobile devices that can be used as a base for many visual recognition tasks. MobileNetV2 is very similar to the original MobileNet, with the exception of the use of inverted residual blocks with bottlenecking features. Lightweight depth-wise convolutions are used as a source of nonlinearity in the intermediate expansion layer to filter features. The architecture of MobileNetV2 includes a 32-filter initial fully convolution layer as well as 19 additional bottleneck layers. It is employed for classification, detection, and alternative common tasks of CNNs that are 53 layers deep.

3.2.3 EfficientNetB7

As a convolution neural network architecture and scaling method EfficientNet [22] scales all the dimensions of depth/width/resolution uniformly using a compound coefficient. By applying compound scaling to the input image, the network can capture more fine-grained patterns on the bigger image by adding more layers and channels. Compared to the previous Gpipe, EfficientNet-B7 achieves 84.4% top-1 / 97.1% top-5 accuracy, and it is 8.4x smaller and 6.1x faster on CPU inference.

3.2.4 DenseNet201

DenseNet-201 [23] is a 201-layer convolutional neural network with feed-forward connections between each layer. DenseNets have several compelling advantages, including the ability to solve the vanishing-gradient problem, improve feature propagation, encourage feature reuse, and significantly reduce the number of parameters. On the majority of them, DenseNets outperform the state-of-the-art while requiring less memory and computation to achieve high performance.

3.2.5 ResNet101

The ResNet [24] achieved outstanding results in the ImageNet and MS-COCO competitions [6]. A convolutional neural network with 101 layers is called ResNet-101. They noticed relative improvements of 28% when ResNet-101 was used to replace VGG-16 layers in Faster R-CNN. Residual Networks, often known as ResNets, is an extremely ingenious architecture that combines the input of every CNN block with its output. The foundation of residual blocks is a connection known as a skip connection.

3.2.6 EfficientNetV2L

Tan and Le [25] have come up again with EfficientNetv2 which have faster training speed and better parameter efficiency compared to previous convolutional neural networks [7] and their main idea is to train on smaller image sizes and test on larger image sizes. Training-aware neural architecture search and scaling are used by the authors to develop these models in order to jointly optimize training speed. As a result of the smaller kernel size of 3x3, EfficientNetV2 adds more layers to compensate for the reduced receptive field. The EfficientNetV2 models expect float tensors of pixels with values between [-1, 1] as inputs in the absence of pre-processing. Similarly, by modifying this architecture by EfficientNetV2L [8] and with the number of classes in our model the structure is defined.

The Pokémon character classification is useful for Pokémon lovers who are eager to know their character names. We developed our model using different pre-trained models for the classification of Pokémon characters. MobileNetV2, EfficientNetB7, EfficientNetV2L, DenseNet201, ResNet101, VGG19, and VGG16 Pre-trained models were used in this work and compared on their performance based on accuracy in classification. Based on the results it showed that ResNet101 will be the best model among the other transfer learning models taken in this work. Thus, this Transfer learning model can automate the classification of Pokémon characters by displaying its name.  Future studies can focus on exploring the performance of transfer learning techniques on other types of visual data, such as videos or 3D images. Additionally, further investigations can be conducted to compare the performance of different transfer learning models on various image classification tasks beyond Pokémon characters.