Kurly Overcomes Deep Learning Limitations... Unveils Laplacian Filter-Based Blurry Image Detection Technology

Kurly's Data Service Development Team revealed their technical journey to effectively detect blurry or dark images among customer purchase reviews. Initially, they utilized a CNN model (implemented in PyTorch) with a structure similar to AlexNet, but due to minimal differences in characteristics between actual review image groups, precision remained at around 0.6.

In an attempt to train on artificially blurred data, validation precision reached 0.9, but due to overfitting per batch, test precision only reached 0.75, failing to meet expectations. Analysis showed that the way CNNs extract local features through filters has limitations in capturing the ambiguity of edges.

Introducing a Rule-Based Algorithm Using the Laplacian Filter

Ultimately, instead of deep learning, they adopted a 'Simple is the best' approach that leverages the structural characteristics of image data. This method converts color images to grayscale, then applies a Laplacian filter to extract edges. The approach is based on the insight that blurry images have unclear edges, resulting in a standard deviation of pixel values that is noticeably smaller than in clear images.

Experimental results showed that when the threshold was set to 3.5, about 90% of the detected images were confirmed to be actually blurry, significantly improving precision. This case demonstrates that a rule-based algorithm can be more efficient than a complex model for a specific problem (blurry image detection).