Log detection method based on weight sharing and upper and lower feature fusion

Abstract

The invention provides a log detection method based on a weight sharing and adjacent upper and lower feature fusion model, and belongs to the field of machine vision and deep learning. According to the method, a pooling feature pyramid is constructed to realize weight sharing features, and feature fusion from top to bottom is simplified into feature fusion of two adjacent layers, so that the complexity of model fusion and the transmission of redundant information are reduced; and the two improved methods enable the precision of a detector to be obviously improved, especially a self-made log data set. Ghe characteristics of dense small targets and data lack are fully considered, the robustness and the training speed of the model are improved through feature fusion and weight sharing, and amore reliable analysis data source is provided in practical application.

Description

technical field [0001] The invention belongs to the field of machine vision and deep learning, relates to deep learning target detection and recognition technology, in particular to a log detection method based on weight sharing and upper and lower feature fusion, and proposes to use the maximum pooling operation to construct a pooled feature pyramid, so that the feature Bottom-up transfer of information flow for feature-sharing log detection method. Background technique [0002] In the traditional SSD algorithm, the feature maps of different layers are independently used as the input of the classification network, so the same object may be detected by frames of different sizes at the same time, and the detection effect on small scales is relatively poor. At present, feature fusion is performed through top-down deconvolution, so that feature information flows from deep layers to shallow prediction modules. BN layer normalization is used before fusion to avoid covering and su...

AI Technical Summary

Problems solved by technology

[0002] In the traditional SSD algorithm, the feature maps of different layers are independently used as the input of the classification network, so the same object may be detected by frames of different sizes at the same time, and the detection effect on small scales is relatively poor.

Images