Image_classification_CIFAR10

INTRODUCTION : Artificial neural networks (ANNs) and training models are the focus of the discipline of deep learning (DL), a branch of machine learning, which is a branch of artificial intelligence. The globe is currently being revolutionised by the ambitious and astonishing field of study known as "deep learning". Emerging technologies that use deep learning technology include Google's DeepDream and self-driving cars. As a result, it is developing into a lucrative subject to learn and work in the twenty-first century. DEEP DREAM SELF DRIVING CARS Image categorization is a famous deep learning application that is quite popular among deep learning scientists. Deep learning makes it possible for photos to be automatically recognised and categorised. The objective of this research is to develop an image classification algorithm utilising the well-known CIFAR-10 dataset. 60,000 photos from ten different classes make up the dataset. The collection also includes 50,000 training images and 10,000 test images. The dataset can still be used as a learning and training tool for convolutional deep learning neural networks for picture categorization even after it has been effectively solved. The CIFAR-10 dataset and the CIFAR-100 collection were developed by researchers at the Canadian Institute for Advanced Research, or CIFAR—the term for Canadian Institute for Advanced Research. The collection includes 60,000 3232 pixel colour images of objects divided into 10 categories, such as frogs, birds, cats, ships, etc. The class names and the corresponding ordinal integer values are listed below. ● 0:aeroplane ● 1:automobile ● 2:bird ● 3:cat ● 4:deer ● 5:dog ● 6:frog ● 7:horse ● 8:ship ● 9:truck The dataset, which consists of very tiny images, was created for computer vision research.

LITERATURE SURVEY : Recognition of objects visually is essential for humans to interact with one other and their surrounding environment. Because we recognise faces, animals, and food almost instantly in daily life, we are able to recognise objects visually. The ventral stream of the human visual system carries out the ability of core object recognition. [1]. It has long been assumed that biological systems are able to recognise visual objects. Many computer vision experiments have made an effort to imitate human item identification. The accuracy of machine visual identification remained much lower than that of a person despite years of study. However, deep learning and convolutional neural networks (CNN) have substantially improved machine performance to the point that it now even surpasses that of humans. [7, 8]. Deep learning-based object recognition enables the quick and precise prediction of an object's location within an image. The object detector uses deep learning, a potent machine learning technique, to automatically acquire the visual characteristics necessary for detection tasks. Similar to standard neural networks, which are modelled after the neural networks found in biological systems, convolutional neural networks for object recognition have a feedforward design and consist of many layers stacked on top of one another. Numerous studies show how Convolutional Neural Networks and those deployed in the human object recognition system are analogous.[9]. Recently, several designs for computer vision, particularly for visual identification based on deep neural networks, have been presented. [5, 8, 19, 20, 21, 22]; although the first convolutional neural network (LeNet) was formulated to recognize characters written by hand [23].Some of the layers that generally make up a CNN are convolution (conv), rectified linear unit (ReLU), batch normalisation (BN), pooling, dropout, and fully connected layer (FC). A block can be created by combining a few fundamental levels. The FC layers are frequently employed at the network's termination. A cost function between the target (actual true value) and network output is reduced in order to find the parameters or weights of the neural network. Optimization almost exclusively employs gradient descent with back propagation. For efficient convergence, a few commencement and weight update techniques have been devised.[7, 26]. Deep neural networks for visual object recognition Sno. ARCHITECTURE NAME

STRUCTURE ACCURACY

1. Alex Net ● There are eight learnable layers in it. ● RGB photos are used as the model's input. ● It has a combination of max-pooling layers on each of its five convolutional layers. ● It then has three completely linked layers. ● Relu is the activation function utilised in all levels. ● Two Dropout layers were utilised. ● Softmax is the activation function utilised in the output layer. ● There are 62.3 million parameters in this architecture as a whole.[34] AlexNet in PyTorch CIFAR10 Class(74.74% Test Accuracy)
2. Google Net ● The auxiliary classifiers' architectural features are as follows: ● a typical 5x5 filter with a three-stride pooling layer. ● a 1-1 convolution with 128 filters for dimension reduction and Rectified linear activation. ● a layer with 1024+1 outputs and Rectified linear activation that is fully connected. ● regularisation of dropouts at a dropout ratio of 0.73. ● a softmax classifier with 1000 classes that produces results similar to the basic softmax classifier. ● [35] GoogleNet in PyTorch CIFAR10 Class(92.57% Test Accuracy)
3. VGG 13 ● Three-by-three filters are used in the first two convolutional layers. With VGG in PyTorch the same convolutions applied to the top two layers' 64 filters, 22422464 volumes are generated. The filters have a 33 step at all times. ● The volume's height and breadth were then decreased from 22422464 to 11211264 using a pooling layer. This layer had a stride of 2 and a max-pool of size 22. This was finished after two more convolution layers with 128 filters. As a result, the new dimension is 112112128. ● When the pooling layer is used, the volume is once more reduced to 5656128. ● The size is decreased to 2828256 after two further convolution layers with 256 filters each are added. ● The 77512 volume is flattened into a Densely Integrated (FC) layer with 4096 channels and a softmax output of 1000 classes after the final pooling layer. CIFAR10 Class(90.17% Test Accuracy)
4. VGG By suggesting a ground-breaking design and improving recognition performance on the ImageNet dataset, VGG rose to become one of the most well-known CNNs. VGG is unique since each convolutional kernel is three by three in size. ReLU is applied to each convolutional layer, and some of them also receive max-pooling. Since VGG uses three FC levels at the very end, the network doesn't drop out. To maintain an equivalent complexity per layer when a feature's size is decreased in half, width (i.e., the quantity of feature maps or filters) must be doubled. Here is another way to interpret such a structure.
5. Inception/GoogLeNet GooLeNet implements the Inception architecture for ImageNet and features 22 layers, a variety of parameters, and several Inception modules. These distinctive branches signify different scales of processing. To match or minimise dimensions, the 11 convolution layer is also employed prior to the 33 or 55 convolution layers (before addition). Actually, the Network-in-Network design was what first proposed the 11 convolutional layers. It can increase representational power while maintaining a consistent receptive field size when used with a conventional convolutional layer.

3. Alexnet Alex Krizhevsky served as AlexNet's chief architect. It is a CNN, or convolutional neural network. AlexNet gained notoriety after competing in the ImageNet Large Scale Visual Recognition Challenge. With a top-5 error of 15.3%, it was successful. This is behind the winning entry by 10.8%. Five convolutional layers, three max-pooling layers, two normalisation layers, two fully connected layers, and one softmax layer make up the AlexNet architecture. Convolutional filters and the ReLU nonlinear activation function make up each convolutional layer. Max pooling is carried out using the pooling layers. Because all of the layers are fully connected, the input size is fixed. Although padding makes the input size appear to be 224x224x3, it is actually 227x227x3. DATASET Alex Krizhevsky served as AlexNet's chief architect. It is a CNN, or convolutional neural network. AlexNet rose to prominence after partaking in the ImageNet Large Scale Visual Recognition Challenge. With a top-5 error of 15.32%, it was successful. This is behind the winning entry by 10.8%. 5 convolutional layers, 3 max-pooling layers, 2 normalisation layers, 2 fully connected layers, and 1 softmax layer make up the AlexNet architecture. Convolutional filters and the ReLU nonlinear activation function make up each fully connected layer. Pooling layers help us carry out the max pooling. Because all of the layers are fully connected, the input size is fixed. Although padding renders the input size seems to be 224x224x3, it is truly 227x227x3.

REQUIREMENTS : LIBRARIES USED :

1. Pytorch
2. Torchvision
3. Matplotlib
4. Numpy
5. Joblib

METHODOLOGY : STEP 1 : LOADING THE DATASET The CIFAR-10 dataset is supplied with the following classes: truck, ship, horse, frog, horse, bird, cat, deer, plane, vehicle, and cat. All of the axes of the photos have been normalised to 0.5. Utilising the torchvision.dataset.CIFAR10, a training and testing set is produced (). With a batch size of 4, the data is put into the train loader and test loader. Using DataLoader from torch.utils.data (). STEP 2 : DEFINING A CONVOLUTIONAL NEURAL NETWORK CNN is a feed-forward network. Error , cost function, are being calculated in the training process, and back propagation updates the weight of the layers. All backward functions compute the gradients of the trainable parameters, whereas the forward function computes the values of the cost function. To build a neural network, utilise the PyTorch.nn package and the torch library. The package consists of extensible classes, modules that are required in the path of formation of a neural network CNN is a neural network which is designed for the detection of complicated features in data. In the implementation 3 CNN layers have been defined, in which all layers follow a hierarchical sequence of cnn, leakyRelu and maxpooling i.e. Conv2d is used with a kernel size of 5 and maxpool with size 2. To chain the layers and the activation functions into a single network architecture we are going to use nn.sequential. The following are the layers that make up our network: ● Numerous features in the input image are extracted by our neural network's first layer, which is one of a set of layers.. Between the layers and a filter we perform convolution that has a specified size of NxN. Then we go over the image we have put in and dot product is being taken between the parts of our input image and the used filter with keeping the size in respect(NxN). The output which we get so far gives us the information about the image, the information regarding the edges and the vertices of the image and is called a Featured map . Then we put this featured map over other layers to explore the other features of the input image ● The activation function here is ReLU which defines all features which got through the layer as 0 or greater than (>0). Which means if we apply this function the negative number becomes zero and the positive ones are kept the same ● Location of an object can affect the ability of out cnn system to detect the features specifically hence, Maxpool layer comes into the picture which helps us ensure no features are being affected. Impact of Using Dropout in PyTorch ● An unregularized network quickly overfits on the training dataset. The validation loss for without-dropout run diverges a lot after just a few epochs. This accounts for the higher generalisation error. ● Training with a dropout layer with a dropout probability of 25% prevents model from overfitting. However, this brings down the training accuracy, which means a regularised network has to be trained longer. ● Dropout improves the model generalisation. Even though the training accuracy is lower than the unregularized network, the overall validation accuracy has improved. This accounts for a lower generalisation error. To achieve regularisation, so that you can Effectively avert the complication of overfitting we use dropout over the layers. In order to make a prediction via neural network the forward method comes into the picture. information is flowing forward through the hidden layers starting from the input to the output because of which the other word for “making prediction” is running the forward pass. Now comes the backward pass which is runned when we compute parameter updates this is called by calling loss.backward(). The information flows from the input layers to the hidden layers and finally to the output layers. To call the forward method the call function in the nn.Module is used, so that when the forward method is called we run model(input). STEP 3 : DEFINING A LOSS FUNCTION Calculated by the loss function is the output's deviation from the target. Through backpropagation in neural networks, the loss function's value is intended to be decreased. After each optimization iteration, the loss function informs us of the behaviour of the model. We determine model accuracy. It displays the percentage of correctly predicted outcomes. We use the Adam Optimizer of the PyTorch neural network package and a loss function based on the Cross-Entropy loss. The degree to which our network's weights are adjusted in relation to loss is controlled by learning rate (lr). It will be set to 0.001. The training moves more slowly the lower the learning rate. Momentum is set to 0.9 and lr is set at 0.001. STEP 4 : TRAINING THE DATA The data is trained with epoch 4 having a batch size of 12500. Forward pass followed by backward propagation is done in each cycle . Later we perform Adam optimization,which is an optimization method , which can greatly improve the performance of the experiment. All layers have an equal learning rate, which is manually adjusted as training progresses. The predictive performance is now defined as the estimated probability of the model in the test set, or overall accuracy. STEP 5: TESTING THE DATA The loss value is printed for every thousand(1k) batches of images for every iteration. Loss value decreases after every loop.Model correctness is distinct from loss value. The test data are used to calculate the model's accuracy, which displays the percentage of correct predictions made. Here, it will reveal how many of the 10,000 test photos our model was able to accurately predict after each iteration.

RESULT ANALYSIS : The following output is generated. The model stands at an accuracy of 70.62%

FUTURE SCOPE AND CONCLUSION: Computer vision applications can make machines more self-sufficient and ensure product quality. If processing systems are quicker than humans, it should be simple to incorporate inline quality controls, such as 100% controls. Damaged components can be replaced or fixed, increasing the productivity of manufacturing operations. Two major issues in agriculture are water scarcity and yield quality. Tracking satellite imaging of the fields can enable the monitoring and data sharing of irrigation. An additional tool for gathering and reporting irrigation is infrared image processing. The results of this research can then be used to inform pre-harvesting decisions concerning whether to harvest or not.A mix of machine learning and image processing algorithms and methodologies can be used to detect the growth of weeds. 2D images are converted into 3D images with a sense of depth using 3D imaging technology. The 3D model is then rendered and colours and textures are added to make it look realistic. Thanks to such 3D imaging and rendering, doctors can see 3D images of her organs in very high quality that have never been seen before. You can use it to perform delicate procedures and make accurate diagnoses. In this study, we classified images using CNN and the CIFAR-10 dataset. The hierarchical sequence of CNN, LeakyRelu, and Max-Pooling is followed by all three of the CNN layers defined in the code sample below. CNNs are effective at removing features from images. As a result, the neural network model's accuracy increased dramatically, from 65% to 70.62%. To further increase accuracy, you can experiment with CNN model hyperparameters. The number of dense layers, the number of hidden units in each dense layer, the number of epochs, the number of convolutional layers, the number of filters in each convolutional layer, and other hyperparameters can all be adjusted.

REFERENCES : [1] J. J. DiCarlo, D. Zoccolan, and N. C. Rust, "How does the brain solve visual object recognition?," Neuron, vol. 73, no. 3, pp. 415-434, 2012. [2] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, pp. 436–444, 2015. [3] J. Schmidhuber, "Deep learning in neural networks: An overview," Neural Networks, vol. 61, pp. 85-117, 2015. [4] Y. Bengio, A. Courville, and P. Vincent, "Representation Learning: A Review and New Perspectives," PAMI, vol. 35, no. 8, pp. 1798-1828, 2013. [5] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet classification with deep convolutional neural networks," in NIPS, 2012, pp. 1097-1105. [6] Y. Bengio, "Learning Deep Architectures for AI," Foundations and Trends in Machine Learning, vol. 2, no. 1, pp. 1-127 , 2009. [7] K. He, X. Zhang, S. Ren, and J. Sun, "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification," in ICCV, 2015. [8] K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition," in CVPR, 2016. [9] R. M. Cichy, A. Khosla, D. Pantazis, A. Torralba, and A. Oliva, "Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence," Scientific Reports, vol. 6, no. 27755, 2016. [20] S. Xie, R. Girshick, P. Dollár, Z. Tu, and K. He, "Aggregated Residual Transformations for Deep Neural Networks," in CVPR, 2017. [21] T. DeVries and G. W. Taylor, "Improved Regularization of Convolutional Neural Networks with Cutout," arXiv, 2017. [22] X. Gastaldi, "Shake-Shake regularization," arXiv, 2017. [23] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradientbased learning applied to document recognition," Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. [24] S. Ioffe and C. Szegedy, "Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift," in ICML, 2015. [25] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, "Dropout: a simple way to prevent neural networks from overfitting," The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929-1958, 2014. [26] D. Kingma and J. Ba, "Adam: A Method for Stochastic Optimization," in International Conference on Learning Representations (ICLR), 2015. [27] K. Simonyan and A. Zisserman, "Very Deep Convolutional Networks for Large-Scale Image Recognition," arXiv, 2014. [28] C. Szegedy et al., "Going deeper with convolutions," in CVPR, 2015. [29] M. Lin, Q. Chen, and S. Yan, "Network in network," CoRR, 2013. [30] A. Krizhevsky, "Learning Multiple Layers of Features from Tiny Images," Technical report, 2009. [31] A. Vedaldi and K. Lenc, "MatConvNet: Convolutional Neural Networks for MATLAB," in ACM international conference on Multimedia, 2015, pp. 689-692. [32] A. Nguyen, J. Yosinski, and J. Clune, "Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognisable Images," in Computer Vision and Pattern Recognition, 2015. [33]Image Processing And Its Future Implications [34]Alex Krizhevsky University of Toronto [email protected] Ilya Sutskever University of Toronto [email protected] Geoffrey E. Hinton University of Toronto [email protected],”ImageNet Classification with Deep Convolutional Neural Networks”, 2012 [35] Deep Learning: GoogLeNet Explained | by Richmond Alake | Towards Data Science