Neural network structure based on multi-picture target recognition

By designing a multi-picture-based target recognition neural network structure, using multiple-picture simultaneous AI learning and training, the problem of the reduction in recognition accuracy of the existing technology during occlusion and lighting changes is solved, and the distinction and accuracy of dynamic and static targets are achieved. Test.

CN118941778AInactive Publication Date: 2024-11-12SHAANXI ZHIHUI ELECTRONIC TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411142644.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing object detection technology has decreased the accuracy of recognition when facing occlusion and lighting changes, and it is difficult to distinguish between dynamic and static targets, especially in the motion of the camera.

Method used

Design a multi-picture target recognition neural network structure, by receiving multi-channel tensor data synthesized by multiple pictures at the input layer, using multiple pictures to simultaneous AI learning and training, including input layer, feature extraction layer and detection layer, to perform Preprocessing of pictures such as grayscale, geometric cropping and filtering, and predict target positions and categories using YOLO, SSD or Faster R-CNN models in the detection layer.

Benefits of technology

It reduces the impact of lighting on target detection, retains the complete information of the target object, can effectively distinguish dynamic and static targets, and improves the accuracy of detection and the ability to analyze video sequences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941778A_ABST
    Figure CN118941778A_ABST
Patent Text Reader

Abstract

The invention discloses a neural network structure based on multi-picture target recognition, which comprises a target detection neural network, the target detection neural network structure is provided with a neural network structure supporting simultaneous input of a plurality of pictures, and an inlet layer (input layer) of the neural network structure receives a picture tensor with a channel number of n * ch, wherein n is the number of pictures which can be received at the same time in the layer, and ch is the number of picture channels. Compared with the prior art, the neural network structure based on multi-picture target recognition has the advantages that the neural network structure based on multi-picture target recognition can be used based on multiple pictures, illumination and shielding sensitivity are reduced, and the detection accuracy is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a neural network structure based on multi-image target recognition. Background Art

[0002] Target detection technology in the field of AI recognition is an important application in the current and future fields of artificial intelligence. The current target detection technologies mainly include target detection algorithms based on deep learning:

[0003] Two-stage Algorithms

[0004] 1. R-CNN series (including R-CNN, Fast R-CNN, Faster R-CNN): Introduces deep learning for feature extraction, with high accuracy but high computational cost, especially R-CNN and Fast R-CNN are inefficient in generating region proposals.

[0005] 2. Mask R-CNN: It adds instance segmentation function based on Faster R-CNN, with high accuracy, but high computational complexity and slow speed.

[0006] One-stage Algorithms

[0007] 1.YOLO (You Only Look Once) series (including YOLOv1 to YOLOv5): It is known for its speed and can achieve real-time detection, but especially in early versions, the detection and positioning accuracy of small targets is not as good as the two-stage algorithm.

[0008] 2.SSD (Single Shot MultiBox Detector): It also seeks a balance between speed and accuracy, but may perform poorly when faced with highly overlapping or small objects.

[0009] 3. RetinaNet: It is designed to solve the problem of imbalance between positive and negative samples in the one-stage algorithm, but its accuracy may be slightly lower than that of some optimized two-stage algorithms.

[0010] The above neural network structure is sensitive to occlusion and lighting changes when used:

[0011] The performance of most algorithms will degrade when the target is occluded or the lighting conditions change drastically. When occlusion occurs or the lighting changes, the texture and contour of the image also change, and the recognition accuracy decreases.

[0012] At the same time, the detection accuracy of dynamic targets is not high:

[0013] Because the existing target detection performs image analysis on each picture in the video instead of analyzing the video sequence, there is no distinction between movement and stillness for the above target detection neural network.

[0014] When the camera is in motion, it is difficult to accurately track dynamic targets:

[0015] Because the existing target detection performs image analysis on each picture in the video instead of analyzing the video sequence, the above algorithms only perceive relatively static images, so they cannot distinguish whether the target is static or dynamic. Summary of the invention

[0016] The technical problem to be solved by the present invention is to overcome the above technical defects and provide a neural network structure based on multi-image target recognition that can be used based on multiple images, reduce sensitivity to illumination and occlusion, and ensure detection accuracy.

[0017] To solve the above technical problems, the technical solution provided by the present invention is: a neural network structure based on multi-image target recognition, including a target detection neural network, the target detection neural network is provided with more channel entrances, the target detection neural network is provided with an entrance design structure for simultaneously inputting multiple images, the structure of the entrance layer of the target detection neural network is designed to receive multi-channel tensor data synthesized by multiple images, and the number of channels accepted by the entrance layer of the neural network structure is n*ch, wherein n is the number of images simultaneously received by the entrance layer, and ch is the number of image channels.

[0018] Preferably, the target detection neural network includes an input layer, a feature extraction layer, a detection layer and an output layer, the input layer is provided with >3 channel entrances, the feature extraction layer includes a convolution layer, a pooling layer and a fusion layer, and the detection layer includes predicting the target position and category.

[0019] Preferably, the preprocessing of the image data includes extracting sequential images from the video stream, including continuous images at equal intervals and discontinuous images at random intervals, and inputting the selected images into an input layer.

[0020] Preferably, the image is further subjected to grayscale processing, geometric cropping, normalization and filtering before being input to the input layer.

[0021] Preferably, the detection layer includes a YOLO, SSD or Faster R-CNN model.

[0022] The advantages of the present invention compared with the prior art are: the present invention adopts a mode and method of simultaneous AI learning and training of multiple pictures, and the neural network no longer relies on the information of a single picture, so the detected information is more accurate. The input of multiple pictures retains the complete normal information of the target object and reduces the impact of light on target detection. The use of multiple pictures and sequence pictures retains the context information and can well distinguish dynamic targets from static targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the neural network structure based on multi-image target recognition. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0025] Combined with Figure 1 As shown, a neural network structure based on multi-image target recognition includes a target detection neural network. The neural network is provided with an input layer design structure for simultaneously inputting multiple images. The structure of the input layer of the neural network is designed to receive multi-channel tensor data synthesized from multiple images. The number of channels accepted by the input layer of the neural network structure is n*ch, where n is the number of images simultaneously received by the entry layer, ch is the number of image channels, and n>1.

[0026] In a specific implementation of the present invention, the target detection neural network includes an input layer, a feature extraction layer, a detection layer and an output layer, wherein the input layer accepts multiple images to be input simultaneously, the feature extraction layer includes a convolution layer, a pooling layer and a fusion layer, the detection layer includes predicting the target position and category, wherein the preprocessing of the image data includes extracting a sequence of images for the video stream, including continuous images at equal intervals and discontinuous images at random intervals, and inputting the selected images into the input layer, and also includes grayscale processing, geometric cropping, normalization and filtering of the images before inputting into the input layer.

[0027] In one embodiment, the detection layer includes a YOLO, SSD, or Faster R-CNN model to predict the location and category of the target.

[0028] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

[0029] The working principle of the present invention is: using an n*3-entry target detection neural network to extract multiple sequence pictures from a video stream, which can also be continuous pictures with equal intervals;

[0030] For example, starting from the current picture, the 1st, 2nd, 3rd, 4th...nth picture, also includes the 1st, 3rd, 5th, 7th, 9th...nth picture, the 1st, 2x, 3x, 4x, 5x...nxth picture.

[0031] After preprocessing, multiple images are converted into n*3 channel data, and after preprocessing, they are input into the input layer neural network with more channels for AI training or prediction results.

[0032] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A neural network structure based on multi-image target recognition, characterized by: It includes a target detection neural network, the neural network structure has an input layer that can accept multiple images as input at the same time, the input layer structure of the target detection neural network is multi-channel tensor data that receives a synthesis of multiple images, and the number of channels accepted by the input layer of the neural network structure is n*ch, where n is the number of images that the neural network input layer can receive simultaneously, ch is the number of channels for each image, and n>1.

2. The neural network structure based on multi-image target recognition according to claim 1, characterized in that: The neural network includes a neural network that can accept multiple images as input at the same time: Including picture image classification neural network, target recognition neural network, and target detection neural network.

3. The neural network structure based on multi-image target recognition according to claim 1, characterized in that: The preprocessing of the image data includes extracting a sequence of images from the video stream, including continuous images at equal intervals and discontinuous images at random intervals, and inputting the selected images into an input layer for use.