Last week I saw this article that took an approach with feature engineering with some color and texture features (+segmentation & gray-level co-occurrence matrix), and then applies a SVM for one-vs-one classification on the PlantVillage dataset. The authors claim to perform not that much worse compared with DL methods. I did not have the time to try this out, but I did wonder whether - even if worse than neural nets - this would be useful as variety in an ensemble. It also made me wonder whether anyone actually had any success with approaches not heavily based on DL. I have to admit that until seeing that article, I did not even seriously consider it as a serious option for image data (particularly when inference time is of limited importance)… Perhaps that was too narrow-minded?!
", "messages": [ { "id": "1209481", "postDate": "02/18/2021 23:24:20", "content": "Last week I saw this article that took an approach with feature engineering with some color and texture features (+segmentation & gray-level co-occurrence matrix), and then applies a SVM for one-vs-one classification on the PlantVillage dataset. The authors claim to perform not that much worse compared with DL methods. I did not have the time to try this out, but I did wonder whether - even if worse than neural nets - this would be useful as variety in an ensemble. It also made me wonder whether anyone actually had any success with approaches not heavily based on DL. I have to admit that until seeing that article, I did not even seriously consider it as a serious option for image data (particularly when inference time is of limited importance)… Perhaps that was too narrow-minded?!
", "rawMarkdown": "Last week I saw [this article](https://arxiv.org/pdf/2102.04515v1.pdf) that took an approach with feature engineering with some color and texture features (+segmentation & gray-level co-occurrence matrix), and then applies a SVM for one-vs-one classification on the PlantVillage dataset. The authors claim to perform not that much worse compared with DL methods. I did not have the time to try this out, but I did wonder whether - even if worse than neural nets - this would be useful as variety in an ensemble. It also made me wonder whether anyone actually had any success with approaches not heavily based on DL. I have to admit that until seeing that article, I did not even seriously consider it as a serious option for image data (particularly when inference time is of limited importance)... Perhaps that was too narrow-minded?!", "votes": null }, { "id": "1209502", "postDate": "02/18/2021 23:48:10", "content": "In year 2011 and before, hand crafted features combined with SVM is what won the annual ImageNet competitions. Then starting in 2012, CNN began winning. (First win was AlexNet in 2012, ResNet in 2015, etc). I'm not sure how much attention hand crafted features plus SVM has received in the last 9 years but it would be interesting to see if they add variety to an ensemble.
", "rawMarkdown": "In year 2011 and before, hand crafted features combined with SVM is what won the annual ImageNet competitions. Then starting in 2012, CNN began winning. (First win was AlexNet in 2012, ResNet in 2015, etc). I'm not sure how much attention hand crafted features plus SVM has received in the last 9 years but it would be interesting to see if they add variety to an ensemble.", "votes": null }, { "id": "1214449", "postDate": "02/22/2021 21:36:08", "content": "I am recently working on some research level in adding GLSZM, GLCM, .. etc other statistical features combined with SVM and it's working well so far on medical diagnosis, no big experience just some non-deep experiments
", "rawMarkdown": "I am recently working on some research level in adding GLSZM, GLCM, .. etc other statistical features combined with SVM and it's working well so far on medical diagnosis, no big experience just some non-deep experiments", "votes": null } ], "comments": [ { "id": 1209502, "author_name": "cdeotte", "author_url": "", "post_date": "02/18/2021 23:48:10", "content": "In year 2011 and before, hand crafted features combined with SVM is what won the annual ImageNet competitions. Then starting in 2012, CNN began winning. (First win was AlexNet in 2012, ResNet in 2015, etc). I'm not sure how much attention hand crafted features plus SVM has received in the last 9 years but it would be interesting to see if they add variety to an ensemble.
", "votes": null, "replies": [] }, { "id": 1214449, "author_name": "omarmohamed22", "author_url": "", "post_date": "02/22/2021 21:36:08", "content": "I am recently working on some research level in adding GLSZM, GLCM, .. etc other statistical features combined with SVM and it's working well so far on medical diagnosis, no big experience just some non-deep experiments
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1209481": "Last week I saw [this article](https://arxiv.org/pdf/2102.04515v1.pdf) that took an approach with feature engineering with some color and texture features (+segmentation & gray-level co-occurrence matrix), and then applies a SVM for one-vs-one classification on the PlantVillage dataset. The authors claim to perform not that much worse compared with DL methods. I did not have the time to try this out, but I did wonder whether - even if worse than neural nets - this would be useful as variety in an ensemble. It also made me wonder whether anyone actually had any success with approaches not heavily based on DL. I have to admit that until seeing that article, I did not even seriously consider it as a serious option for image data (particularly when inference time is of limited importance)... Perhaps that was too narrow-minded?!", "1209502": "In year 2011 and before, hand crafted features combined with SVM is what won the annual ImageNet competitions. Then starting in 2012, CNN began winning. (First win was AlexNet in 2012, ResNet in 2015, etc). I'm not sure how much attention hand crafted features plus SVM has received in the last 9 years but it would be interesting to see if they add variety to an ensemble.", "1214449": "I am recently working on some research level in adding GLSZM, GLCM, .. etc other statistical features combined with SVM and it's working well so far on medical diagnosis, no big experience just some non-deep experiments" }, "source": "meta" }