Introduction We want to implement a DL model to classify fine art images by genre, as proposed in the paper linked here: https://ieeexplore.ieee.org/abstract/document/8675906. We think that this will be a meaningful challenge because rather than classifying something with a hard ground truth, such as object recognition, our model seeks to categorize images semantically by artistic style.
Challenges The data preprocessing portion involved a lot more steps than our class assignments. Although we were able to find an existing dataset, its images were not consistently sized, ranging between 117 x 249 and 3000 x 2530. In order to use a sequential model, we needed to resize the images (ideally to 600x600, which is close to the average height / width across the dataset). For images smaller than the target size, this meant padding; for images larger than the target size, we had to shrink them. In addition, this size is much larger than the images we have been using in class— MNIST images are around 28x28. We are able to run our model with a target size of 100x100, but we are still looking for a workaround (perhaps using gcp credits to train) so we can increase the image sizes.
Insights At 12 categories, a randomly guessing model would have an accuracy of 8% ish. Currently, with a target size of 100x100, we have an accuracy of roughly 32%. This is much lower than our target accuracy of 50, but hopefully, by adding additional max pool layers and finding a feasible way to feed larger images into our model, we can increase performance.
Plan Our foundational model is set, and runs successfully. We need to dedicate more time to addressing the issues described under Insights.
Log in or sign up for Devpost to join the conversation.