Film Defect Detection Using Multi-Layer Recognition Model

The detection process is segmented into three distinct functional layers: convolution layer for feature extraction, regression layer for position and size prediction, and classification layer for defect type identification. This segmentation allows each layer to specialize in specific tasks, improving overall detection accuracy while maintaining manageable process complexity through modular architecture.

The patent transforms the detection problem from traditional 2D image analysis to a multi-dimensional approach by incorporating bounding box coordinates (x, y, width, height) and angle information as additional dimensions. This dimensional expansion enables the model to capture irregular shapes and uncertain distribution positions more effectively, significantly improving detection accuracy for complex scratch patterns.

The patent extracts only the essential features needed for scratch detection through the convolution layer, focusing on key characteristics such as edge patterns, orientation, and spatial distribution. By extracting only relevant features rather than processing entire images exhaustively, the system achieves faster detection speed with reduced computational resource consumption.

The model dynamically adjusts detection parameters including bounding box coordinates, angle, and size based on the input image characteristics. This parameter adaptation allows the system to optimize detection speed for different scratch types and densities, improving overall productivity while consuming fewer resources by avoiding uniform high-computation processing for all cases.