Image Target Localization Using Classifier Heatmaps and Unsupervised Detection

The patent combines traditional target detection algorithms with deep learning-based image classification algorithms. The determination module integrates both approaches by first obtaining traditional detection results and then using image classification to determine thermodynamic diagrams, which are processed to generate final positioning frames. This merging resolves the contradiction by incorporating deep learning's superior recognition capabilities while maintaining the structural framework of traditional detection methods.

The patent implements a self-service approach where the system uses its own generated thermodynamic diagrams and positioning frames to improve detection accuracy without requiring extensive external labeled data. The image classification module processes image data to create thermodynamic diagrams that guide the target detection process, reducing dependence on manually labeled training data and thereby reducing labor costs associated with data acquisition and labeling.

The patent segments the target detection process into distinct functional modules: traditional target detection algorithm execution, image classification for thermodynamic diagram generation, and non-maximum suppression for final result optimization. By segmenting these tasks and processing them in sequence rather than simultaneously, the system achieves high accuracy through deep learning while managing computation time through structured processing stages.