Method for judging defective products

Abstract

This invention discloses a method for judging defective products, comprising the steps of: a) training an object detection model and a segmentation model respectively using multiple images of qualified products and images of defective products related to a product, so that in any other image containing the product, the object detection model can find the possible defect locations of the product in the image, and the segmentation model can find that any possible defect location contains a specific pixel representing the defect, thereby judging whether the product contained in the image is a qualified product or a defective product; b) inputting the detection image containing the product into the object detection model, finding the possible defect locations of the product in the detection image, and then cutting out a possible defect image containing a portion of the product's pixels from the detection image based on the possible defect locations; c) inputting the possible defect image into the segmentation model to confirm whether there are specific pixels representing the defect to judge whether the product is a qualified product or a defective product; and d) outputting the judgment result of the segmentation model.

Claims

1. A method for judging defective products, executed by a processor, comprising the steps of: a) training an object detection model and a segmentation model respectively using multiple images of qualified products and images of defective products related to a product, such that in any other image containing the product, the object detection model can find the possible defect locations of possible defects of the product in the image, and the segmentation model can find that any possible defect location contains a specific pixel representing a defect, thereby judging whether the product contained in the image is a qualified product or a defective product; b) inputting a detection image containing the product into the object detection model, finding the possible defect locations of the product in the detection image, and then cutting out a possible defect image containing a portion of the product's pixels from the detection image based on the possible defect locations; c) inputting the possible defect image into the segmentation model, confirming whether there are specific pixels representing defects to judge whether the product is a qualified product or a defective product; and d) outputting the judgment result of the segmentation model.

2. The defective product determination method as described in claim 1, wherein the possible defect location is a matrix of multiple pixel coordinates surrounding the pixels in the product that may have defects.

3. The defective product judgment method as described in claim 1 further includes a step a1) between step a) and step b): cropping pixels within a specific range from the original inspection images to form an inspection image, wherein the applicable original inspection images have the same size, and the product pixels of the product in each original inspection image do not exceed the specific range.

4. The defective product judgment method as described in claim 1, wherein step c) further includes: inputting the detection image into an optical character recognition model capable of recognizing text information contained in the image to obtain product information about the product, wherein step d) further includes: outputting the product information together with the judgment result.

5. The defective product identification method as described in claim 4, wherein the product information is a serial number, product description text, manufacturer, or product size.

6. The defective product judgment method as described in claim 1 further includes a step a2) between step a) and step b): inputting the original inspection image into an anomaly detection model to determine whether the original inspection image contains the product; if the product is contained, the original inspection image is output as the detection image; otherwise, it is discarded and not output, wherein step a) further includes: training the anomaly detection model with multiple different images of the product, so that the anomaly detection model can determine whether any image contains the product.

7. The defective product judgment method as described in claim 1 further includes a step a3) between step a) and step b): inputting the original inspection image into a classification model to determine whether the original inspection image contains the product and the specific object; if it contains the product and the specific object, then the original inspection image is segmented into the inspection image containing the product and the object image containing the specific object, and output separately; if it only contains the product, then the original inspection image is output as the inspection image, wherein step a) further includes: training the classification model with multiple different images of the product and the specific object, so that the classification model can determine whether any image contains the product and / or the specific object.

8. The defective product judgment method as described in claim 1, further comprising a step e) after step d): if the judgment result is a defective product, inputting the potentially defective image into a classification model to determine the type of defect in the product image, and outputting a re-inspection result, wherein step a) further comprises: training the classification model with images of multiple types of defective products for a product, so that the classification model can determine the type of defect of the product and the pixels containing the defect in any image containing the product.

9. The defective product judgment method as described in claim 1, wherein step b) further comprises: flipping, rotating and / or moving the detection images to enhance the judgment effect of each model.

10. The defective product identification method as claimed in claim 1, wherein the possible defect location is one of the following types: when the inspection image contains multiple products to be inspected, the possible defect location is set to include all of the products with possible defects; the possible defect location is set to the periphery of the possible defect, varying with the appearance of the possible defect; or the possible defect location is set to a square image range including the possible defect.

Images