# CNN 捲積 ![](https://i.imgur.com/K9NhhGn.jpg) ![](https://i.imgur.com/DvRKz2R.png) ![](https://i.imgur.com/fX2bQ3m.png) 池化 ![](https://i.imgur.com/iX3d2Bp.jpg) ![](https://i.imgur.com/pd2WIs7.png) yolact ![](https://i.imgur.com/d0htHV2.png) FPN C1做捲積(降採樣) 視覺感受變大 P3-->P4做反捲積(升採樣) 解析度變高 C3跟C4做疊合，得到更多的細節 ![](https://i.imgur.com/76NR1QP.png) 左邊 捲積 降採樣 右邊 反捲積 升採樣 ![](https://i.imgur.com/jhLapxI.png) ![](https://i.imgur.com/cWOkLrZ.png) Anchor ![](https://i.imgur.com/UZaOBBQ.png) resulting in the scales [24, 48, 96, 192, 384] ![](https://i.imgur.com/NAyUfg0.png) x nms ![](https://i.imgur.com/fRI7WaJ.png) pi是幾層P構成的 左邊是原來的retinaNet 本來只輸出class跟 box cordinate anchor aspect ratio(a) | class | box | | -------- | -------- | | class&score | box cordinate&confidence score | yolact | class | box | mask | | ------------ | -------- | -------- | | class&score&anchor| box cordinate&confidence score&anchor | k mask coefficent | ![](https://i.imgur.com/dn9WO2y.png) protonet會產生很多的prototype 之後在做prototype跟mask coefficient 係數相乘 crop 做切割 threshold 邊緣檢測後找出找出完整樣本 2天 斜率檢測製圖演算法開發 4天 水平雜訊過濾演算法開發 3天 垂直雜訊過濾演算法開發 5天 特徵檢測演算法開發 5天 水平座標排序法開發 2天 樣本驗證與演算法最佳化調整 7天 影片合成 1天