Micro-defect Detection Based on Multi-scale Feature Backtracking Structure

A Method for Extracting Planar Image Features Based on a Multi-Scale Feature Backtracking Network in Surface Machining

“…The core idea and method of this thesis are put forward by Haotian Song. "Since during the detection process, utilizing multi-scale features for repeated detection at various scales is essential, this will inevitably slow down the speed of the detection module [10]. In order to accelerate network detection, we have proposed two methods to enhance FPN [13]: (1) decreasing the quantity of data that the feature needs to process; (2) expand the scope of receptive field coverage.…”

“…Compared to single-layer feature detectors, multi-level feature detectors can better handle complex image data and more accurately detect and recognize target objects. Two methods exist for multi-level feature detection: image pyramid-based and feature pyramid-based [10]. The feature pyramid-based strategy is presently the most frequently utilized method in CNN models [3].…”

“…We introduced a discriminant layer in the features of different layers. Suitable experiments have revealed that when ascertaining the current feature value, the network structure can trace back to the value of the correspondent feature in the relative position of the lower layer and relearn using the convolution function [10]. This method can not only effectively simplify network calculations and improve the efficiency of network operation, but also better integrate feature information at different levels, making the final output features more targeted.…”

“…We employ a combination approach, specifically the anchor-based center prior method, to enable the application of feature operations to different targets using weights [9]. At the same time, in order to improve the efficiency and accuracy of data processing, for different targets, we use a Gaussianlike weighted function G to distinguish them, defining d as the bias, and  and  as hyperparameters [10]. The definition of G is:…”

“…Because objects of the identical type have varying sizes and aspect ratios, distinct downsampling rates of different layer features can also cause some interference. In order to permit features at diverse levels to effectively represent target information at assorted scales, initially, the upper layer features are expanded [10], then the stage subscript rate is calculated based on the layer and normalized accordingly.…”