COMSATS University Islamabad (Lahore Campus) □Sessional-1 □ Sessional-II √ Terminal Examination Course Title: Computer Vision Course Instructor/s: Dr. Zeeshan Gillani Semester: 6th Batch: FA18 Section: Time Allowed: 90 Minutes Student’s Name: SHAHROZE NADEEM Important Instructions / Guidelines: a. Start with prayer. b. 50 Mins to solve the paper+10 Mins to Upload = 60 Mins c. Attempt your own exam – Spring 2020 Course Code: CSC462 Credit 3(2,1) Hours: Programme BCS Name: C Date: 40 Maximum Marks: Reg. FA18-BCS-147-C No. Question no 1: Feature selection is integral part in image classification problem. You are required to extract feature descriptor using chain codes for the following example using 4 directional chain codes in anti-clockwise direction. . [10] Question no 2: Perceptron is basic neural network algorithm. Design an architecture of perceptron to solve image classification problem by dividing images into two groups and sample input shown in figure below. If number of brighter pixels are greater than darker pixels than image is classified as brighter image otherwise darker image. [10] Identify number of class Write an activation function Define input feature and size Suitable learning rate Draw a perceptron architecture Question no 3: You are required to calculate the total learnable parameters for following CNN network. In order to improve this network, suggest possible advantages and disadvantages if we opt for more deeper CNN. [10] INPUT CONV1(f=6, s=1, Nc=10) POOL1(f=2, s=2) CONV1(f=6, s=1, Nc=20) POOL2(f=2, s=2) FC3 (200) FC4 (100) Softmax(10) Activation shape 32,32,3 30,30,10 15,15,10 12,12,20 6,6,20 200,1 100,1 10,1 Activation Size 3072 9000 2250 1440 720 120 100 10 Parameters 0 (3*3*3+1)*10=280 0 (3*3*10+1)*20=1820 0 720*120+120=86520 120*100+100=12100 100*10+10=1010 Question no 4: One hot encoded technique was used to transform images into 1 dimension vector and pass to SVM and KNN classifiers for image classification task. Discuss drawback of this technique and suggest more suitable method with example. [5] In one hot encoded technique we face difficulties because of feature space blow up in no time. So that the representation size grows with the corpus. Each vector in one hot encoded is equidistant from every vectors. It is not suitable for tasks such as entity recognition. On contrary a distributed representation is quite much less in size than original input. It captures salient features among different elements of input that’s why one hot encoded is not recommended Question no 5: Describe how CNN uses feature sharing and spare connections to solve complex problem in computer vision and compare them to traditional feature extraction techniques [5] For image recognition, a convolutional neural network is used. It's divided into two sections: convolution and pooling. It's a little easier to train than traditional image recognition methods. It solves complex problems by reducing the number of connections and using feature sharing and spare connections to improve generalization. It's also possible to improve problem solving by replacing the linear filter with a non-linear function.