final model: U-Net ResNet-34 + FPN EfficientNet-B3 (val 0.7297)

README.md

@@ -1,11 +1,11 @@
 ---
 license: mit
 library_name: pytorch
-tags: [image-segmentation, defect-detection, segmentation-models-pytorch]
+tags: [image-segmentation, defect-detection, segmentation-models-pytorch, onnx]
 ---
-# SurfaceGuard — surface defect segmentation
+# SurfaceGuard: surface defect segmentation
 
-Segments four defect types (`defect_a`–`defect_d`) on 128×800 grayscale strips. Final option: **U-Net ResNet-34 + FPN EfficientNet-B3**,
+Segments four defect types (`defect_a` to `defect_d`) on 128×800 grayscale strips. Final option: **U-Net ResNet-34 + FPN EfficientNet-B3**,
 flip TTA, per-class post-processing tuned on a held-out validation fold (2,279 images).
 
 | | |
@@ -24,6 +24,6 @@ flip TTA, per-class post-processing tuned on a held-out validation fold (2,279 i
 | U-Net ResNet-34 | 0.6859 | 0.7084 | 0.6165 | 0.9261 | 0.9155 | 24.4326 |
 | Scratch U-Net | 0.6259 | 0.6759 | 0.5976 | 0.8571 | 0.8620 | 7.8512 |
 
-Files: `weights/*.pt` (state dict + config), `pp_params.json` (threshold, component filter, minimum area,
-presence-gate threshold per class), `selection.json`. Input: grayscale image scaled to [0, 1], shape 1×128×800.
+Files: `weights/*.pt` (PyTorch state dict + config), `pp_params.json` (threshold, component filter, minimum area,
+presence-gate threshold per class), `selection.json`. Input: grayscale image scaled to [0, 1], shape 1×1×128×800.
 Course project, IIT Madras BS Deep Learning & GenAI (Atul Kumar, 24f2003045). Demo: https://huggingface.co/spaces/Sawarn/surfaceguard-demo

model_comparison.csv

@@ -1,8 +1,8 @@
 option,n_models,params_M,ms_per_img,raw_score,score,fg_dice,fp_score,empty_acc,fp_pairs,missed_pairs,dice_defect_a,dice_defect_b,dice_defect_c,dice_defect_d
-Scratch U-Net + U-Net ResNet-34 + FPN EfficientNet-B3,3,44.75976,38.00101351350899,0.7098836680185818,0.730579551974095,0.6632068067140616,0.8861386138613861,0.8910706450197455,874,119,0.4793001398767564,0.442010434775577,0.6912996026502313,0.7522796208883983
-U-Net ResNet-34 + FPN EfficientNet-B3,2,36.908584,23.671430681503985,0.7121077177562986,0.7297188830586218,0.6577928687168699,0.8965607087024492,0.8995173321632295,794,122,0.46703500974354656,0.4342841375934587,0.6855978818846239,0.7523187427847984
-FPN EfficientNet-B3,1,12.476,15.700430449502099,0.6882240430974451,0.7280215163310845,0.6560928872857535,0.8947368421052632,0.8980912681000439,808,121,0.4029141426131736,0.4579558682381054,0.6930091498942201,0.7636085835575438
-Scratch U-Net + FPN EfficientNet-B3,2,20.327176,30.030013281507102,0.6965028024288047,0.7270537224298317,0.6444490966864553,0.9210526315789473,0.9172882843352348,606,148,0.4034437180005399,0.46459504033198973,0.6804383868653375,0.7431614311535297
-Scratch U-Net + U-Net ResNet-34,2,32.28376,22.30058306400689,0.6914241651540225,0.7108353975963919,0.6118631255771155,0.9467170401250652,0.9318780166739798,409,212,0.3790188253291136,0.44152027028238205,0.6414637590495214,0.7190581572515924
-U-Net ResNet-34,1,24.432584,7.971000232001887,0.6859102215467197,0.708351529030656,0.616530765159067,0.9261334028139656,0.9155331285651601,567,203,0.36109615509804116,0.3178738646050214,0.6583635882510604,0.7319145066517528
-Scratch U-Net,1,7.851176,14.329582832005004,0.625918255677087,0.675921840094213,0.5975776391971999,0.8570870244919229,0.862000877577885,1097,161,0.37406755459065866,0.4005747413523921,0.6327049775594353,0.6904011381590637
+Scratch U-Net + U-Net ResNet-34 + FPN EfficientNet-B3,3,44.75976,38.51242868876238,0.7098836680185818,0.730579551974095,0.6632068067140616,0.8861386138613861,0.8910706450197455,874,119,0.4793001398767564,0.442010434775577,0.6912996026502313,0.7522796208883983
+U-Net ResNet-34 + FPN EfficientNet-B3,2,36.908584,23.967867516906388,0.7121077177562986,0.7297188830586218,0.6577928687168699,0.8965607087024492,0.8995173321632295,794,122,0.46703500974354656,0.4342841375934587,0.6855978818846239,0.7523187427847984
+FPN EfficientNet-B3,1,12.476,15.723389562361593,0.6882240430974451,0.7280215163310845,0.6560928872857535,0.8947368421052632,0.8980912681000439,808,121,0.4029141426131736,0.4579558682381054,0.6930091498942201,0.7636085835575438
+Scratch U-Net + FPN EfficientNet-B3,2,20.327176,30.267950734217578,0.6965028024288047,0.7270537224298317,0.6444490966864553,0.9210526315789473,0.9172882843352348,606,148,0.4034437180005399,0.46459504033198973,0.6804383868653375,0.7431614311535297
+Scratch U-Net + U-Net ResNet-34,2,32.28376,22.78903912640078,0.6914241651540225,0.7108353975963919,0.6118631255771155,0.9467170401250652,0.9318780166739798,409,212,0.3790188253291136,0.44152027028238205,0.6414637590495214,0.7190581572515924
+U-Net ResNet-34,1,24.432584,8.244477954544797,0.6859102215467197,0.708351529030656,0.616530765159067,0.9261334028139656,0.9155331285651601,567,203,0.36109615509804116,0.3178738646050214,0.6583635882510604,0.7319145066517528
+Scratch U-Net,1,7.851176,14.544561171855985,0.625918255677087,0.675921840094213,0.5975776391971999,0.8570870244919229,0.862000877577885,1097,161,0.37406755459065866,0.4005747413523921,0.6327049775594353,0.6904011381590637