Adversarial X-Ray Image Decomposition for Cargo-Vehicle Separation
Solution Overview
Problem
Existing image recognition systems struggle to accurately distinguish cargo from vehicle structures in X-ray images due to the overlap of features in existing technologies, and existing technologies fail to efficiently address the acquisition of sufficient training data, which limits the accuracy of cargo inspection systems, and the need for labor-intensive inspection methods, leading to reduced efficiency and increased risk of missed detections.
Innovation Solution
A method and apparatus for training an image decomposition model using a first and second adversarial neural network, each comprising a generation and discrimination network, to decompose X-ray images by adjusting parameters through backpropagation based on loss function values derived from fused training images, enhancing the accuracy of cargo and vehicle separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine intelligent recognition is used to improve inspection efficiency, then processing speed increases, but recognition accuracy deteriorates due to feature overlap in complex X-ray images
Solution Approach 1:
The patent applies segmentation by decomposing the complex X-ray image into multiple component images (cargo image, vehicle body image, and background image) using an image decomposition model. This separates the overlapping cargo features from vehicle structure features, enabling accurate cargo detection while maintaining high processing speed through automated decomposition.
2Measurement precision
If traditional manual image discrimination is used to maintain recognition accuracy, then detection precision improves, but inspection efficiency deteriorates due to complex vehicle structures and uneven cargo stacking
Solution Approach 1:
The patent introduces an image decomposition model as an intermediary between the raw X-ray image and the cargo detection process. This model automatically separates cargo features from vehicle structure features, providing clean decomposed cargo images to the detection algorithm, thereby achieving high accuracy without manual intervention.
3Measurement precision
If supervised learning with large amounts of labeled data is used to improve recognition accuracy, then model performance improves, but data acquisition workload and time increase significantly
Solution Approach 1:
The patent uses a generative adversarial network (GAN) to generate synthetic labeled images as copies of real scenarios. The GAN creates training images with automatic labels by simulating cargo and vehicle structures, eliminating the need for manual labeling while providing sufficient training data for supervised learning.
Solution Approach 2:
The system performs self-labeling through the image decomposition model and GAN framework. The model automatically generates labels for training images by identifying and separating cargo regions from vehicle structures, enabling the system to create its own training data without external human annotation.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method of training an image decomposition model, an apparatus of training an image decomposition model, a method of decomposing an image, an electronic device, and a storage medium are provided, which may be applied to a field of an image processing technology. The method of training the image decomposition model includes: acquiring a training set; inputting a first training image into the first adversarial neural network and inputting a second training image into the second adversarial neural network, so as to determine a first loss function value; inputting a third training image into the first adversarial neural network and the second adversarial neural network respectively, so as to determine a second loss function value; determining a third loss function value according to a comparison result between an acquired fusion image and the third training image, where the fusion image is generated by fusing a generated image of the first adversarial neural network with a generated image of the second adversarial neural network; and adjusting a parameter of the image decomposition model according to at least one of the first to third loss function values, until a training end condition is met, where the first loss function, the second loss function and the third loss function meet a predetermined weight relationship.