Attention-Weighted Image Quality Evaluation for Target Recognition

The patent segments the image processing task by dividing images into multiple sub-regions and evaluating each sub-region separately. The quality evaluation network model processes different sub-regions independently, extracting features from each region and combining them to form a comprehensive quality assessment. This segmentation allows the system to handle variations in posture, size, and occlusion by focusing on relevant sub-regions while ignoring irrelevant areas.

The patent applies local quality evaluation by assigning different weights to different sub-regions based on their importance for target recognition. The attention mechanism calculates attention weights for each sub-region, giving higher weights to regions containing the target object and lower weights to background regions. This local quality approach enables the system to maintain high evaluation accuracy even when overall image quality varies.

The patent extracts only the relevant features needed for target recognition by using the quality evaluation network model to identify and focus on sub-regions containing the target object. The attention mechanism extracts features from specific sub-regions rather than processing the entire image, effectively taking out the useful target features while leaving out the interfering non-target features. This selective extraction improves identification accuracy by eliminating irrelevant information.

Instead of extracting features from the whole image and then filtering out non-target features, the patent inverts the approach by first identifying relevant sub-regions and then extracting features only from those regions. The quality evaluation network model evaluates sub-region importance and guides feature extraction accordingly, extracting features from important sub-regions while ignoring unimportant areas. This inverted approach prevents the inclusion of non-target features in the first place.

The patent performs preliminary action by pre-training the quality evaluation network model with a large dataset of images with various qualities, postures, and occlusion levels. The model learns to recognize quality patterns and features during the training phase, enabling it to make accurate quality assessments during actual processing without requiring complex real-time computations. This preliminary training equips the model with the knowledge needed for accurate evaluation while maintaining efficient processing speed.

The patent introduces an attention mechanism as an intermediary between the input images and the quality evaluation process. The attention mechanism computes attention weights for different sub-regions and uses these weights to guide the feature extraction and quality assessment. This intermediary layer enables the model to focus computational resources on the most relevant parts of the image, improving evaluation accuracy without requiring excessive processing power across the entire image.