Multi-layer aggregation for object detection

Abstract

Object detection (58) uses a deep or multiple layer network (72-80) to learn features for detecting (58) the object in the image. Multiple features from different layers are aggregated (46) to train (48) a classifier for the object. In addition or as an alternative to feature aggregation from different layers, an initial layer (72) may have separate learnt nodes for different regions of the image (70) to reduce the number of free parameters. The object detection (58) is learned or a learned object detector is applied.

Description

Background technique [0001] This embodiment relates to object detection and machine learning for object detection. [0002] For machine-learned object detection, input features from images are used to train and apply a detector. The quality of features is critical to the performance of many image analysis tasks. Scientists have come up with various features leveraging their deep understanding of the data and tasks at hand. For example, Haar features are used in organ detection and segmentation due to their cheap computation. Local binary pattern (LBP) features are good for representing shape or texture and are suitable for human re-identification. No feature is optimal for all tasks or all types of data. Often a considerable amount of experience is required to select good features for a particular application. [0003] Evolution of sensing technology or changes in existing sensing technology occasionally necessitate the design of new features. This processing is often ch...

AI Technical Summary

Problems solved by technology

While it is possible to fine tune the network using discriminative training, the output may not capture all relevant information

This can be caused by missing labeled samples or by backpropagation optimization that suffers from bad local minima

Furthermore, the fully connected structure of deep networks makes it difficult to learn with large images or volumes

Images