CN-121981941-A - Semiconductor wafer core particle defect detection method and device based on improvement YOLOv8

CN121981941ACN 121981941 ACN121981941 ACN 121981941ACN-121981941-A

Abstract

The invention discloses a semiconductor wafer core particle defect detection method and device based on an improvement YOLOv, and the method and device comprise the steps of constructing a semiconductor wafer core particle defect detection model based on an improvement YOLOv and training to obtain a trained semiconductor wafer core particle defect detection model, replacing a neck network from a PAN-FPN network to a shortcut pyramid network on the basis of a YOLOv model by the semiconductor wafer core particle defect detection model, adding an attention network between the backbone network and the neck network, respectively pasting semiconductor wafer core particles to be detected in the back bottom of Lan Zhengguang, blue backlight or red positive light to obtain a wafer core particle image, and inputting the wafer core particle image into the trained semiconductor wafer core particle defect detection model to obtain defect types, coordinates and confidence in the semiconductor wafer core particles to be detected. The invention solves the problem of low detection precision of the wafer core particle defects.

Inventors

CAI WENBI
LIU ZE
GAO YUE
DING GUIGUANG

Assignees

泉州市三安集成电路有限公司
清华大学

Dates

Publication Date: 20260505
Application Date: 20251208

Claims (10)

1. The semiconductor wafer core particle defect detection method based on the improvement YOLOv is characterized by comprising the following steps of: constructing and training a semiconductor wafer core particle defect detection model based on an improvement YOLOv to obtain a trained semiconductor wafer core particle defect detection model, replacing a neck network from a PAN-FPN network to a shortcut pyramid network on the basis of a YOLOv model by the semiconductor wafer core particle defect detection model, and adding an attention network between a backbone network and the neck network; and respectively pasting the semiconductor wafer core particles to be detected in the back bottoms of Lan Zhengguang, blue backlight or red positive light, shooting to obtain wafer core particle images, and inputting the wafer core particle images into the trained semiconductor wafer core particle defect detection model to obtain defect types, coordinates and confidence in the semiconductor wafer core particles to be detected.
2. The semiconductor wafer core defect detection method based on the improvement YOLOv according to claim 1, wherein the attention network includes CBAM module, the semiconductor wafer core defect detection model retains the output characteristics of the second C2f module and the output characteristics of the SPPF module in the backbone network on the basis of the YOLOv model, and makes them input into the shortcut pyramid network for feature fusion after passing through the CBAM module, and deletes the detection head for outputting the maximum size feature in the header network.
3. The method for detecting a semiconductor wafer core defect based on claim YOLOv as defined in claim 2, wherein the shortcut pyramid network includes a first splice layer, a first C2f module, an upsampling module, a second splice layer, a second C2f module, a third splice layer, a third C2f module, a downsampling module, a fourth splice layer, and a fourth C2f module; the output characteristics of a second C2f module in a backbone network are obtained after being subjected to CBAM module and the output characteristics of an SPPF module are respectively used as a first input characteristic and a second input characteristic of the shortcut pyramid network, the first input characteristic and the second input characteristic are input into the first splicing layer to obtain a first splicing characteristic, the first splicing characteristic is input into a first C2f module of the shortcut pyramid network to obtain a first fusion characteristic, the first fusion characteristic is input into the second splicing layer after being subjected to the up-sampling module and the first input characteristic to obtain a second splicing characteristic, the second splicing characteristic is input into a second C2f module of the shortcut pyramid network to obtain a second fusion characteristic, the second fusion characteristic and the first input characteristic are input into the third splicing layer to obtain a third splicing characteristic, the third splicing characteristic is input into the third C2f module of the shortcut pyramid network to obtain a first fusion characteristic, the fourth fusion characteristic is input into the fourth splicing layer after being subjected to the up-sampling module and the fourth input characteristic is input into the fourth splicing layer, and obtaining a fourth fusion feature, wherein the third fusion feature and the fourth fusion feature are used as output features of different scales of the shortcut pyramid network.
4. The method for detecting semiconductor wafer core particle defects based on the improvement YOLOv as defined in claim 1, wherein the semiconductor wafer core particle defect detection model is quantified using a quantization-aware technique during training of the semiconductor wafer core particle defect detection model, the quantization-aware technique used including maximum symmetric static quantization.
5. The method for detecting defects of semiconductor wafer core particles based on the improvement YOLOv as set forth in claim 4, wherein the training process of the semiconductor wafer core particle defect detection model includes: constructing a semiconductor wafer core particle defect image dataset; performing preliminary training on the semiconductor wafer core particle defect detection model by utilizing the semiconductor wafer core particle defect image dataset to obtain a semiconductor wafer core particle defect detection model after preliminary training; And carrying out quantization sensitivity analysis on the primarily trained semiconductor wafer core particle defect detection model, screening out a quantization sensitive layer and a quantization insensitive layer in the primarily trained semiconductor wafer core particle defect detection model, keeping the weight and bias in the quantization sensitive layer in fp16 format, carrying out int8 quantization on the weight and bias in the quantization insensitive layer, and carrying out fine adjustment on the semiconductor wafer core particle defect detection model by utilizing the semiconductor wafer core particle defect image data set after quantization to obtain the trained semiconductor wafer core particle defect detection model.
6. The method of claim 1, wherein the defect classification comprises at least one of color non-uniformity, grain smudge, edge chipping, epitaxial ring shape, electrode smudge, finger electrode breakage, finger electrode scratch, ITO loss, poly, PN ring loss, and electrode scratch.
7. A semiconductor wafer core particle defect detection apparatus based on improvement YOLOv, comprising: A model building module configured to build and train a semiconductor wafer core particle defect detection model based on the improvement YOLOv, resulting in a trained semiconductor wafer core particle defect detection model that replaces the neck network from the PAN-FPN network to a shortcut pyramid network on the basis of the YOLOv model, and adds an attention network between the backbone network and the neck network; The prediction module is configured to paste semiconductor wafer core particles to be detected in the back of Lan Zhengguang, blue backlight or red positive light respectively to obtain wafer core particle images, and input the wafer core particle images into the trained semiconductor wafer core particle defect detection model to obtain defect types, coordinates and confidence in the semiconductor wafer core particles to be detected.
8. An electronic device, comprising: One or more processors; storage means for storing one or more programs, When executed by the one or more processors, causes the one or more processors to implement the method of any of claims 1-6.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-6.
10. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any of claims 1-6.

Description

Semiconductor wafer core particle defect detection method and device based on improvement YOLOv8 Technical Field The invention relates to the field of defect detection, in particular to a semiconductor wafer core particle defect detection method and device based on improvement YOLOv. Background In modern semiconductor manufacturing processes, wafer Die (Die) inspection is an important component of product quality control, and the inspection accuracy and efficiency are directly related to the production yield and the enterprise cost. During the process of manufacturing, dicing and packaging, many types of defects such as scratches, electrode smudges, electrode defects, color irregularities, etc. often occur in the wafer core. The defect features are weak and the boundaries are fuzzy, so that the traditional detection means are extremely challenged. Traditional methods such as image algorithms based on template matching or manual feature extraction, while effective in certain regular defects, perform poorly for complex background, multiple defect types, low contrast images, and are difficult to accommodate for the high-throughput production line requirements. The wafer core particle defect detection has the following difficulties: 1) Defects tend to be tiny targets (e.g., scratches, dirt, polycrystals, etc.); 2) The defects of the same kind have various forms, directions and textures; 3) The defect is not clearly contrasted with the background. YOLO (You Only Look Once) series of algorithms are widely used for industrial defect detection due to the characteristics of end-to-end, high speed and high precision. However, the YOLO model still has some drawbacks when dealing with complex background and multi-scale targets, such as fine scratch defects on the core particle and slight chromatic aberration, which are difficult to accurately detect. Disclosure of Invention The application aims to provide a semiconductor wafer core particle defect detection method and device based on improvement YOLOv for solving the technical problems. In a first aspect, the present invention provides a semiconductor wafer core particle defect detection method based on improvement YOLOv, comprising the steps of: constructing and training a semiconductor wafer core particle defect detection model based on an improvement YOLOv to obtain a trained semiconductor wafer core particle defect detection model, replacing a neck network from a PAN-FPN network to a shortcut pyramid network on the basis of a YOLOv model by the semiconductor wafer core particle defect detection model, and adding an attention network between a backbone network and the neck network; and respectively pasting the semiconductor wafer core particles to be detected in the back bottoms of Lan Zhengguang, blue backlight or red positive light, shooting to obtain wafer core particle images, and inputting the wafer core particle images into a trained semiconductor wafer core particle defect detection model to obtain defect types, coordinates and confidence in the semiconductor wafer core particles to be detected. Preferably, the attention network comprises CBAM modules, the semiconductor wafer core particle defect detection model reserves the output characteristics of a second C2f module and the output characteristics of the SPPF module in the backbone network on the basis of the YOLOv model, and enables the output characteristics to be input into the shortcut pyramid network after passing through CBAM modules for characteristic fusion, and a detection head used for outputting the maximum size characteristics in the head network is deleted. The rapid pyramid network comprises a first splicing layer, a first C2f module, an up-sampling module, a second splicing layer, a second C2f module, a third splicing layer, a third C2f module, a down-sampling module, a fourth splicing layer and a fourth C2f module, wherein the output characteristics of the second C2f module in the backbone network are obtained through the CBAM module and the output characteristics of the SPPF module and are respectively used as a first input characteristic and a second input characteristic of the rapid pyramid network, the first input characteristic and the second input characteristic are input into the first splicing layer to obtain a first splicing characteristic, the first splicing characteristic is input into the first C2f module of the rapid pyramid network to obtain a first fusion characteristic, the first fusion characteristic is input into the second splicing layer with the first input characteristic after passing through the up-sampling module, the second splicing characteristic is input into the second C2f module of the rapid network to obtain a second fusion characteristic, the output characteristic is obtained through the CBAM module and is respectively used as the first input characteristic and the second input characteristic of the rapid pyramid network, the first input characteristic and the